Skip to content

Automation foundation model for tiny devices: 2-bit, 8-29 MB, tool calls, ASR, structured extraction and embeddings on phones, wearables, smart homes, robots, cars and microcontrollers.

Needle 3

HenryNdubuakud835fb487a1a
diff --git a/README.md b/README.md
index e93064f..d0da12c 100644
--- a/README.md
+++ b/README.md
@@ -1,44 +1,36 @@
-![Needle](assets/banner.png)
+![Needle](assets/banner.svg)
-# Needle 3
+[![Try it in the browser](https://img.shields.io/badge/Try%20it%20in%20the%20browser-cactuscompute.com%2Fneedle-e8593d?style=flat-square)](https://cactuscompute.com/needle)
+[![Hugging Face](https://img.shields.io/badge/Hugging%20Face-Cactus--Compute%2Fneedle3-ffd21e?style=flat-square&logo=huggingface&logoColor=000)](https://huggingface.co/Cactus-Compute/needle3)
+[![GitHub](https://img.shields.io/badge/GitHub-cactus--compute%2Fneedle-181717?style=flat-square&logo=github)](https://github.com/cactus-compute/needle)
-Needle 3 is a 25-121M parameter foundation model for tool calling and structured extraction on tiny devices. It is a Laddered Simple Attention Network: any depth from 2 to 20 layers is a complete model, compressed to CQ2-bit with Cactus Quants into one 9-35 MB weights file that a sub-1 MB engine loads at start.
+# Needle
-This repository is the Python package: inference, LoRA fine-tuning, and export. `pip install cactus-needle`, describe your tools, and call them from Python. The engine and the weights are fetched once from Hugging Face and cached; there is nothing else to build.
+A foundation model for mobiles, wearables, robots, smart home, automotive and microcontrollers. The whole model is a single 8-29 MB binary built on our Simple Attention Network, and we trade general chat capacity to beat models 10x its size on mobile tool calls and match 2-3x bigger models on extraction.
-- **Self-contained**: a sub-1 MB engine per platform plus one `needle3.cact` weights file; inference does no network.
-- **Simple contract**: tool calls come back as structured data, text in, JSON out; a byte-level grammar compiled from your schemas constrains every token.
-- **Confidence-gated**: every response carries a calibrated confidence score from a learned head; set a threshold, act above it, escalate below it.
-- **Bounded memory**: a 256-token sliding window with the tools pinned as KV sinks, so memory stays flat no matter how long the conversation runs.
+- **Tool calls**: given the functions your app exposes, Needle picks the right ones and fills every argument from what the user said. Ask for two things and you get two calls in order; ask for something no tool covers and you get an empty list, not a guess.
+- **Structured extraction**: declare a shape, hand over messy text, get typed fields back: an invoice, a booking, a notification, a form. The decode grammar guarantees the output parses, and extraction generalises to classification.
+- **Text embedding**: the same model returns a vector for a sentence, so an app can search, match and route locally.
-Weights: [huggingface.co/Cactus-Compute/needle3](https://huggingface.co/Cactus-Compute/needle3) · source: [github.com/cactus-compute/needle](https://github.com/cactus-compute/needle).
+![Needle 3 at a glance](assets/model.svg)
-![Size-quality frontier: mobile-class and below](assets/frontier.png)
+Needle 3 is a Laddered Simple Attention Network: a Monarch Hadamard MLP in place of the FFN, GQA attention with causal conv taps, engram n-gram memory read by gather, and multi-lane hyper-connections, trained so that every depth from 2 to 20 layers is a deployable model. Most of its parameters sit in the engram, so the 121M model does the arithmetic of a 50M one. A byte-level grammar compiled from your schemas constrains every token, and every response carries a calibrated confidence score from a learned head. The architecture diagram is on the [release page](https://cactuscompute.com/needle).
-## Laddered Simple Attention Network
+## Benchmarks
-Needle 3 is a Laddered Simple Attention Network, our small-model recipe: a Monarch Hadamard MLP in place of the FFN, GQA attention with causal conv taps, engram n-gram memory read by gather, and multi-lane hyper-connections, trained so that every depth from 2 to 20 layers is a deployable model. See the paper for the design and ablations: [arXiv:2607.18363](https://arxiv.org/abs/2607.18363).
+Tool calling is exact-match accuracy on the full test splits, extraction is field micro-F1 on the full test splits.
-![Simple Attention Network architecture](assets/architecture.png)
+![Needle 3 against baselines on six benchmarks](assets/benchmarks.svg)
-Each block carries its update rule. Here x̂ is the RMS-normalised flattening of the four residual streams, H the orthonormal Walsh-Hadamard transform (a fixed matrix, applied in n log n time with no weights to read), (kₜ, vₜ) rows gathered from hashed n-gram tables, and P the doubly-stochastic normalisation of the routing logits A, computed by Sinkhorn iteration; a, b, g and all σ-gates are learned and input-dependent. Both attention and MLP residuals are sandwich-normed and gated, the engram sites fire at two layers, and decoding is constrained by a byte-level grammar compiled from the declared schemas.
+The interactive frontier plot, the architecture and the fine-tuning results are at [cactuscompute.com/needle](https://cactuscompute.com/needle).
-## Quickstart
+## Get started
```sh
pip install cactus-needle
```
-The runtime package does not install the training stack. Add the `train` extra
-when using fine-tuning or checkpoint export:
-
-```sh
-pip install "cactus-needle[train]"
-```
-
-Needle reads your tool descriptions to decide what to call and how to fill arguments, so describing them well is the whole game.
-
-**Simple**: decorate a function. The signature gives the argument types, the docstring is the tool description, and `run()` completes the loop: model picks the call, Needle executes your function, feeds the result back, and returns the final response with the executed tool results attached as `results`.
+Decorate a function: the signature gives the argument types, the docstring is the tool description, and `run()` completes the loop, executing your function and returning its results.
```python
import needle
@@ -53,127 +45,59 @@ print(agent.run("what's it like in Lagos right now?")["results"])
# [{'city': 'Lagos', 'temp_c': 27, 'sky': 'clear'}]
```
-**Extraction**: to pull structured data out of text, declare the shape and call `extract()`. Pass a Pydantic model and you get a typed object back.
-
-```python
-from pydantic import BaseModel
-
-class Invoice(BaseModel):
- vendor: str
- total: float
- due_date: str
-
-invoice = needle.extract("Invoice from Acme Corp, $1,200.00, due 2026-09-01", Invoice)
-print(invoice.vendor, invoice.total) # -> Acme Corp 1200.0
-```
-
-Per argument descriptions and choices, value constraints compiled into the decode grammar, raw JSON schemas, driving the loop with `complete()`, the response contract, system facts, tool retrieval, and confidence gating are all covered in [doc/apis.md](doc/apis.md).
-
-## Playground
-
-Try any model in the browser: pick a preset, edit the tools or prompt, and Run. Follow-up queries continue the same conversation.
-
-```sh
-needle playground # base model, http://127.0.0.1:7860
-needle playground --weights my.cact # a tuned model
-```
-
-The server downloads and initializes the model before serving, so the first query is instant. The **Finetune on these tools** button runs the fine-tuning pipeline below from the UI and hands back a downloadable `.cact`.
-
-## Environments
-
-Ready-made tool surfaces in `needle.environments`: `smart_home`, `media_player`, `productivity`, `wearable`, `kitchen_appliance`, and `data_capture`. Each is a hand-curated set of tools whose enums, bounds, and descriptions map cleanly onto Needle's constrained decoding, with a ready agent and a frozen acceptance suite.
-
-```python
-from needle.environments import smart_home
-
-smart_home.agent.complete("dim the study lights to 30 percent")
-smart_home.run_tests()
-```
-
-`python -m needle.environments.smart_home` runs a suite from the shell. To adapt an environment to your product, swap the `Literal` values (rooms, contacts, categories) for your own and keep the shapes: closed sets as enums, bounded numbers, verbatim copy for free text, five tools or fewer. The full tool surfaces and the suite contract are in [doc/environments.md](doc/environments.md).
-
-## Fine-tuning
-
-Needle fine-tunes with LoRA on the frozen base and merges the adapter at export, so a run is cheap and the tuned model is still a single `.cact` that runs on the same engine. The workflow is: (optionally) synthesize data, LoRA fine-tune, then build a tuned `.cact`. See [doc/finetuning.md](doc/finetuning.md) for dataset sizing, reading the loss curve, and troubleshooting.
-
-**Data format.** A JSONL file, one example per line. `reasoning` is optional; an off-topic example has `answers: []`.
-
-```json
-{"query": "dim the kitchen to 10", "tools": [{"name": "set_lights", "parameters": {"type": "object", "properties": {"room": {"type": "string"}, "brightness": {"type": "integer"}}, "required": ["room"]}}], "answers": [{"name": "set_lights", "arguments": {"room": "kitchen", "brightness": 10}}], "reasoning": "'kitchen' -> room; 'dim to 10' -> brightness 10"}
-```
-
-**1. Synthesize data (optional).** Needs `OPENROUTER_API_KEY`. Seed from a tool schema file, or expand an existing set:
+Every turn returns one JSON object with `function_calls`, the model's `reasoning` and a calibrated `confidence`; an off-topic request returns an empty list rather than a guess. `needle.Needle(tools=[...], generation=2)` keeps running Needle 2 for existing deployments.
-```sh
-export OPENROUTER_API_KEY=sk-or-...
-needle generate-data --tools my_tools.json --num-samples 500 --output data.jsonl
-needle generate-data --augment data.jsonl --num-samples 500 # expand an existing JSONL
-```
+## Guides
-Set `OPENROUTER_URL` to use an OpenAI-compatible gateway instead of the default OpenRouter endpoint.
+- [How to design tools for Needle 3](https://cactuscompute.com/blog/designing-tools-for-needle): one tool per action, names users would say, formats in descriptions, constraints in the grammar, triggers.
+- [Leveraging Needle's confidence](https://cactuscompute.com/blog/needle-confidence): what the score measures, what the engine withholds, and routing on act, confirm or refuse.
+- [Structured JSON extraction with Needle](https://cactuscompute.com/blog/structured-extraction-with-needle): the record as the only tool, typed results, classification with enums.
+- [Fine-tuning Needle](https://cactuscompute.com/blog/finetuning-needle): the data format, the commands, reading the loss, sizing the dataset.
+- [Needle Python docs](https://cactuscompute.com/blog/needle-python-docs): the API, the response shape, the behaviour contract, system facts, tool retrieval, offline devices, environments, the CLI.
+- [What devices are supported on Needle](https://cactuscompute.com/blog/needle-supported-devices): every platform folder, the CLI runner, the C API, the browser, WASI, air-gapped setup.
+- [The .cact format](https://cactuscompute.com/blog/cact-format): the file the engine maps and reads in place, Cactus Quants at 2.125 bits per weight, and how to parse it yourself.
-**2. LoRA fine-tune.** The base checkpoint auto-downloads from Hugging Face if you do not pass `--checkpoint`. `--generate N` first synthesizes N more examples from the tools in your data (also needs `OPENROUTER_API_KEY`).
+`llms.txt` in this repo carries the same reference for AI coding assistants.
-```sh
-needle finetune data.jsonl --epochs 10
-needle finetune data.jsonl --epochs 10 --generate 300 --lora-rank 16 --lora-alpha 32
-```
+## Customisation
-Key options: `--epochs` (default 3), `--lora-rank` (16), `--lora-alpha` (32), `--lr` (1e-4), `--batch-size` (16), `--max-len` (1024), `--val-split` (0.1), `--checkpoint <base.safetensors or .pkl>`, `--checkpoint-dir <dir>` (default `checkpoints`), `--out <adapter.safetensors or .pkl>`, `--generate <n>`, `--model <id>` (default `deepseek/deepseek-v4-flash`), and `--workers <n>` (default 8). `--generate` uses the configured OpenRouter endpoint to synthesize extra examples before training. The adapter is written to `checkpoints/needle_lora.pkl` by default. A validation loss prints each epoch from the held out split.
+Needle was designed to be customised. Its capacity is a ladder, and a subnetwork as small as 2 layers, fine-tuned on one product's tools, runs optimally on devices far smaller than the full model needs. Fine-tuning on DroidCall lifts every subnetwork by 18 to 36 points, and from 4 layers up the tuned subnetwork passes DeepSeek V4 Flash, starting at 29M parameters.
-Training is plain JAX and runs on any accelerator jax supports. On an NVIDIA machine install the CUDA build and the same command trains on the GPU:
+![Every subnetwork before and after fine-tuning on DroidCall and on Mobile Actions](assets/finetune.svg)
```sh
-pip install "cactus-needle[train,gpu]"
+pip install "cactus-needle[train]"
+needle finetune data.jsonl --epochs 10 --out adapter.safetensors
+needle build --lora adapter.safetensors --layers 8 --out tuned.cact
```
-On Apple Silicon the `metal` extra trains on the GPU:
+Local fine-tuning trains and exports at 4 bits; the [fine-tuning guide](https://cactuscompute.com/blog/finetuning-needle) has the rest. The 2-bit post-training and quantisation behind the shipped model, enriched with Cactus proprietary datasets, run on the [Cactus Platform](https://cactuscompute.com/dashboard).
-```sh
-pip install "cactus-needle[train,metal]"
-```
+## Deploy
-Fine-tuning trains LoRA through 4-bit quantisation-aware numerics at the full 20 layers, so the adapter matches the archive `needle build` writes.
+Every deployment target ships a prebuilt engine under 1 MB that loads the `needle3.cact` weights at start. `needle build --platform <folder> [--layers N]` fetches that engine and puts the weights beside it.
-**3. Build a tuned `.cact`.** Merge the adapter into the base and quantise to 4 bits. The checkpoint is optional: `needle build` uses the base the adapter was trained on, and auto-downloads the Needle 3 base if absent. Every build fetches the published `needle3.cact` and takes its tokenizer from it.
+![One engine per platform folder](assets/deploy.svg)
```sh
-needle build --lora checkpoints/needle_lora.safetensors --out my_needle.cact
-needle build --lora checkpoints/needle_lora.safetensors --layers 8 --out my_needle_8l.cact # 8-layer rung of the tuned base
-```
-
-The engine is weights-agnostic and never rebuilt: one engine library per platform (under 1MB) runs any archive. `needle download needle3` fetches the base 20-layer archive by itself, `needle download needle3.safetensors` the checkpoint to fine-tune. `needle build --layers n` exports any rung from 2 to 20 layers of the tuned base.
-
-Set `NEEDLE_HF_REPO=<you>/<model>` and pass `--upload` to publish the `.cact`. The counterpart `needle download <you>/<model>/my_needle.cact` pulls a published archive on any machine, and `needle download <platform>` (e.g. `macos-arm64`) fetches that platform's engine runner.
-
-**4. Run it.** The engine is weights-agnostic, so a tuned `.cact` runs on it directly - no recompilation:
-
-```python
-import needle
-agent = needle.Needle(weights="my_needle.cact", tools=[...])
-agent = needle.Needle(tools=[...]) # the base Needle 3 archive, fetched once and cached
-agent = needle.Needle(tools=[...], generation=2) # Needle 2, for existing deployments
-agent.run("...")
+needle build --platform macos-arm64
+needle build --platform linux-arm64 --layers 8 --out ./pi
+./macos-arm64/needle --model needle3.cact --tools tools.json --serve
```
-## Telemetry
-
-Cactus Compute collects strictly anonymous usage telemetry (function name, package version, OS — never your prompts, outputs, or data); opt out with `NEEDLE_TELEMETRY=0` or `DO_NOT_TRACK=1`.
+The [devices guide](https://cactuscompute.com/blog/needle-supported-devices) lists every folder and what ships in it.
## Citation
-Needle 3 is built by the Cactus Compute team. If you use it in your work, please cite:
+Needle is built by the Cactus Compute team. If you use it in your work, please cite:
```bibtex
@misc{needle3_2026,
- title = {Needle 3: A Foundation Tool-Calling Model for Tiny Devices},
+ title = {Needle: Automation Foundation Model for Tiny Devices},
author = {Ndubuaku, Henry and Mosoyan, Karen and Mroz, Jakub and Cylich, Noah and
Kumar, Satyajit and Sandhu, Parkirat and Shemet, Roman and Lee, Justin H.},
year = {2026},
organization = {Cactus Compute, Inc.},
howpublished = {\url{https://github.com/cactus-compute/needle}}
}
-```
-
-Reach out on [email protected] for partnerships, collaborations, synergies and deploying Needle2 in your product.
+```
\ No newline at end of file
diff --git a/assets/architecture.png b/assets/architecture.png
deleted file mode 100644
index d27c08e..0000000
Binary files a/assets/architecture.png and /dev/null differ
diff --git a/assets/banner.png b/assets/banner.png
deleted file mode 100644
index ce1940d..0000000
Binary files a/assets/banner.png and /dev/null differ
diff --git a/assets/banner.svg b/assets/banner.svg
new file mode 100644
index 0000000..01d3341
--- /dev/null
+++ b/assets/banner.svg
@@ -0,0 +1,12 @@
+<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 1200 480" width="1200" height="480" role="img" aria-label="Cactus Needle 3">
+<style>svg{--t0:0.0s}.s{animation-duration:9.0s;animation-iteration-count:infinite;animation-fill-mode:both;animation-delay:calc(0s - var(--t0))}@keyframes draw{0%{stroke-dashoffset:1;opacity:1}2.22%{stroke-dashoffset:1}28.89%{stroke-dashoffset:0}93.33%{stroke-dashoffset:0;opacity:1}98.89%{opacity:0}100%{opacity:0;stroke-dashoffset:1}}.draw{animation-name:draw;animation-timing-function:cubic-bezier(.4,0,.2,1)}@keyframes word{0%{width:0}17.78%{width:0}32.22%{width:720px}93.33%{width:720px}98.89%{width:0}100%{width:0}}.word{animation-name:word;animation-timing-function:steps(13,end)}@keyframes three{0%{opacity:0;transform:translateY(12px)}33.33%{opacity:0;transform:translateY(12px)}37.78%{opacity:1;transform:translateY(0)}93.33%{opacity:1;transform:translateY(0)}98.89%{opacity:0}100%{opacity:0}}.three{animation-name:three;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes needle{0%{transform:scaleX(0);opacity:1}35.56%{transform:scaleX(0)}44.44%{transform:scaleX(1)}93.33%{transform:scaleX(1);opacity:1}98.89%{opacity:0}100%{opacity:0;transform:scaleX(1)}}.needle{animation-name:needle;transform-box:fill-box;transform-origin:left center;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes tag{0%{opacity:0}42.22%{opacity:0}47.78%{opacity:1}93.33%{opacity:1}98.89%{opacity:0}100%{opacity:0}}.tag{animation-name:tag}@media (prefers-reduced-motion:reduce){.s{animation:none!important}.draw{stroke-dashoffset:0}.word{width:720px}.three,.tag{opacity:1;transform:none}.needle{transform:scaleX(1)}}</style>
+<defs><clipPath id="cw"><rect class="s word" x="430" y="150" width="0" height="130"/></clipPath></defs>
+<rect width="1200" height="480" fill="#0e0e10"/>
+<g transform="translate(64 72) scale(0.37)">
+<path class="s draw" d="M 272 848 L 272 545 C 272 534 262 530 222 530 A 190 190 0 0 1 32 340 L 32 170 A 64 64 0 0 1 160 170 L 160 340 C 160 372 187 399 219 399 L 272 399 L 272 207 A 175 175 0 0 1 623 207 L 623 620 C 623 645 640 655 665 655 L 736 655 L 736 432 A 64 64 0 0 1 864 432 L 864 640 A 163 163 0 0 1 701 803 L 655 803 C 636 803 623 815 623 835 L 623 848" fill="none" stroke="#ffffff" stroke-width="64" stroke-linejoin="round" stroke-linecap="butt" pathLength="1" stroke-dasharray="1"/>
+</g>
+<text x="430" y="248" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="86" font-weight="600" fill="#ffffff" letter-spacing="-2.5" clip-path="url(#cw)">Cactus Needle</text>
+<text class="s three" x="1024" y="248" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="86" font-weight="600" fill="#e8593d" letter-spacing="-2.5">3</text>
+<rect class="s needle" x="430" y="274" width="690" height="3" rx="1.5" fill="#e8593d"/>
+<text class="s tag" x="430" y="322" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="18" letter-spacing="1.5" fill="#85827a">AUTOMATION FOUNDATION MODEL FOR TINY DEVICES</text>
+</svg>
\ No newline at end of file
diff --git a/assets/benchmarks.svg b/assets/benchmarks.svg
new file mode 100644
index 0000000..1219f96
--- /dev/null
+++ b/assets/benchmarks.svg
@@ -0,0 +1,336 @@
+<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 900 482" width="900" height="482" role="img" aria-label="Needle 3 against baselines on six benchmarks, one benchmark at a time">
+<style>svg{--t0:0.0s}.s{animation-duration:20.4s;animation-iteration-count:infinite;animation-fill-mode:both;animation-delay:calc(0s - var(--t0))}@keyframes sc0{0%{opacity:0}0.00%{opacity:0}1.47%{opacity:1}15.20%{opacity:1}16.67%{opacity:0}100%{opacity:0}}.sc0{animation-name:sc0}@keyframes b0_0{0%{transform:scaleX(0)}1.23%{transform:scaleX(0)}4.66%{transform:scaleX(1)}15.20%{transform:scaleX(1)}100%{transform:scaleX(0)}}.b0_0{animation-name:b0_0;transform-box:fill-box;transform-origin:left center;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes v0_0{0%{opacity:0}3.68%{opacity:0}5.15%{opacity:1}15.20%{opacity:1}100%{opacity:0}}.v0_0{animation-name:v0_0}@keyframes b0_1{0%{transform:scaleX(0)}1.52%{transform:scaleX(0)}4.95%{transform:scaleX(1)}15.20%{transform:scaleX(1)}100%{transform:scaleX(0)}}.b0_1{animation-name:b0_1;transform-box:fill-box;transform-origin:left center;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes v0_1{0%{opacity:0}3.97%{opacity:0}5.44%{opacity:1}15.20%{opacity:1}100%{opacity:0}}.v0_1{animation-name:v0_1}@keyframes b0_2{0%{transform:scaleX(0)}1.81%{transform:scaleX(0)}5.25%{transform:scaleX(1)}15.20%{transform:scaleX(1)}100%{transform:scaleX(0)}}.b0_2{animation-name:b0_2;transform-box:fill-box;transform-origin:left center;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes v0_2{0%{opacity:0}4.26%{opacity:0}5.74%{opacity:1}15.20%{opacity:1}100%{opacity:0}}.v0_2{animation-name:v0_2}@keyframes b0_3{0%{transform:scaleX(0)}2.11%{transform:scaleX(0)}5.54%{transform:scaleX(1)}15.20%{transform:scaleX(1)}100%{transform:scaleX(0)}}.b0_3{animation-name:b0_3;transform-box:fill-box;transform-origin:left center;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes v0_3{0%{opacity:0}4.56%{opacity:0}6.03%{opacity:1}15.20%{opacity:1}100%{opacity:0}}.v0_3{animation-name:v0_3}@keyframes b0_4{0%{transform:scaleX(0)}2.40%{transform:scaleX(0)}5.83%{transform:scaleX(1)}15.20%{transform:scaleX(1)}100%{transform:scaleX(0)}}.b0_4{animation-name:b0_4;transform-box:fill-box;transform-origin:left center;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes v0_4{0%{opacity:0}4.85%{opacity:0}6.32%{opacity:1}15.20%{opacity:1}100%{opacity:0}}.v0_4{animation-name:v0_4}@keyframes b0_5{0%{transform:scaleX(0)}2.70%{transform:scaleX(0)}6.13%{transform:scaleX(1)}15.20%{transform:scaleX(1)}100%{transform:scaleX(0)}}.b0_5{animation-name:b0_5;transform-box:fill-box;transform-origin:left center;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes v0_5{0%{opacity:0}5.15%{opacity:0}6.62%{opacity:1}15.20%{opacity:1}100%{opacity:0}}.v0_5{animation-name:v0_5}@keyframes b0_6{0%{transform:scaleX(0)}2.99%{transform:scaleX(0)}6.42%{transform:scaleX(1)}15.20%{transform:scaleX(1)}100%{transform:scaleX(0)}}.b0_6{animation-name:b0_6;transform-box:fill-box;transform-origin:left center;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes v0_6{0%{opacity:0}5.44%{opacity:0}6.91%{opacity:1}15.20%{opacity:1}100%{opacity:0}}.v0_6{animation-name:v0_6}@keyframes b0_7{0%{transform:scaleX(0)}3.28%{transform:scaleX(0)}6.72%{transform:scaleX(1)}15.20%{transform:scaleX(1)}100%{transform:scaleX(0)}}.b0_7{animation-name:b0_7;transform-box:fill-box;transform-origin:left center;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes v0_7{0%{opacity:0}5.74%{opacity:0}7.21%{opacity:1}15.20%{opacity:1}100%{opacity:0}}.v0_7{animation-name:v0_7}@keyframes b0_8{0%{transform:scaleX(0)}3.58%{transform:scaleX(0)}7.01%{transform:scaleX(1)}15.20%{transform:scaleX(1)}100%{transform:scaleX(0)}}.b0_8{animation-name:b0_8;transform-box:fill-box;transform-origin:left center;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes v0_8{0%{opacity:0}6.03%{opacity:0}7.50%{opacity:1}15.20%{opacity:1}100%{opacity:0}}.v0_8{animation-name:v0_8}@keyframes b0_9{0%{transform:scaleX(0)}3.87%{transform:scaleX(0)}7.30%{transform:scaleX(1)}15.20%{transform:scaleX(1)}100%{transform:scaleX(0)}}.b0_9{animation-name:b0_9;transform-box:fill-box;transform-origin:left center;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes v0_9{0%{opacity:0}6.32%{opacity:0}7.79%{opacity:1}15.20%{opacity:1}100%{opacity:0}}.v0_9{animation-name:v0_9}@keyframes b0_10{0%{transform:scaleX(0)}4.17%{transform:scaleX(0)}7.60%{transform:scaleX(1)}15.20%{transform:scaleX(1)}100%{transform:scaleX(0)}}.b0_10{animation-name:b0_10;transform-box:fill-box;transform-origin:left center;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes v0_10{0%{opacity:0}6.62%{opacity:0}8.09%{opacity:1}15.20%{opacity:1}100%{opacity:0}}.v0_10{animation-name:v0_10}@keyframes b0_11{0%{transform:scaleX(0)}4.46%{transform:scaleX(0)}7.89%{transform:scaleX(1)}15.20%{transform:scaleX(1)}100%{transform:scaleX(0)}}.b0_11{animation-name:b0_11;transform-box:fill-box;transform-origin:left center;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes v0_11{0%{opacity:0}6.91%{opacity:0}8.38%{opacity:1}15.20%{opacity:1}100%{opacity:0}}.v0_11{animation-name:v0_11}@keyframes sc1{0%{opacity:0}16.67%{opacity:0}18.14%{opacity:1}31.86%{opacity:1}33.33%{opacity:0}100%{opacity:0}}.sc1{animation-name:sc1}@keyframes b1_0{0%{transform:scaleX(0)}17.89%{transform:scaleX(0)}21.32%{transform:scaleX(1)}31.86%{transform:scaleX(1)}100%{transform:scaleX(0)}}.b1_0{animation-name:b1_0;transform-box:fill-box;transform-origin:left center;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes v1_0{0%{opacity:0}20.34%{opacity:0}21.81%{opacity:1}31.86%{opacity:1}100%{opacity:0}}.v1_0{animation-name:v1_0}@keyframes b1_1{0%{transform:scaleX(0)}18.19%{transform:scaleX(0)}21.62%{transform:scaleX(1)}31.86%{transform:scaleX(1)}100%{transform:scaleX(0)}}.b1_1{animation-name:b1_1;transform-box:fill-box;transform-origin:left center;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes v1_1{0%{opacity:0}20.64%{opacity:0}22.11%{opacity:1}31.86%{opacity:1}100%{opacity:0}}.v1_1{animation-name:v1_1}@keyframes b1_2{0%{transform:scaleX(0)}18.48%{transform:scaleX(0)}21.91%{transform:scaleX(1)}31.86%{transform:scaleX(1)}100%{transform:scaleX(0)}}.b1_2{animation-name:b1_2;transform-box:fill-box;transform-origin:left center;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes v1_2{0%{opacity:0}20.93%{opacity:0}22.40%{opacity:1}31.86%{opacity:1}100%{opacity:0}}.v1_2{animation-name:v1_2}@keyframes b1_3{0%{transform:scaleX(0)}18.77%{transform:scaleX(0)}22.21%{transform:scaleX(1)}31.86%{transform:scaleX(1)}100%{transform:scaleX(0)}}.b1_3{animation-name:b1_3;transform-box:fill-box;transform-origin:left center;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes v1_3{0%{opacity:0}21.23%{opacity:0}22.70%{opacity:1}31.86%{opacity:1}100%{opacity:0}}.v1_3{animation-name:v1_3}@keyframes b1_4{0%{transform:scaleX(0)}19.07%{transform:scaleX(0)}22.50%{transform:scaleX(1)}31.86%{transform:scaleX(1)}100%{transform:scaleX(0)}}.b1_4{animation-name:b1_4;transform-box:fill-box;transform-origin:left center;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes v1_4{0%{opacity:0}21.52%{opacity:0}22.99%{opacity:1}31.86%{opacity:1}100%{opacity:0}}.v1_4{animation-name:v1_4}@keyframes b1_5{0%{transform:scaleX(0)}19.36%{transform:scaleX(0)}22.79%{transform:scaleX(1)}31.86%{transform:scaleX(1)}100%{transform:scaleX(0)}}.b1_5{animation-name:b1_5;transform-box:fill-box;transform-origin:left center;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes v1_5{0%{opacity:0}21.81%{opacity:0}23.28%{opacity:1}31.86%{opacity:1}100%{opacity:0}}.v1_5{animation-name:v1_5}@keyframes b1_6{0%{transform:scaleX(0)}19.66%{transform:scaleX(0)}23.09%{transform:scaleX(1)}31.86%{transform:scaleX(1)}100%{transform:scaleX(0)}}.b1_6{animation-name:b1_6;transform-box:fill-box;transform-origin:left center;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes v1_6{0%{opacity:0}22.11%{opacity:0}23.58%{opacity:1}31.86%{opacity:1}100%{opacity:0}}.v1_6{animation-name:v1_6}@keyframes b1_7{0%{transform:scaleX(0)}19.95%{transform:scaleX(0)}23.38%{transform:scaleX(1)}31.86%{transform:scaleX(1)}100%{transform:scaleX(0)}}.b1_7{animation-name:b1_7;transform-box:fill-box;transform-origin:left center;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes v1_7{0%{opacity:0}22.40%{opacity:0}23.87%{opacity:1}31.86%{opacity:1}100%{opacity:0}}.v1_7{animation-name:v1_7}@keyframes b1_8{0%{transform:scaleX(0)}20.25%{transform:scaleX(0)}23.68%{transform:scaleX(1)}31.86%{transform:scaleX(1)}100%{transform:scaleX(0)}}.b1_8{animation-name:b1_8;transform-box:fill-box;transform-origin:left center;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes v1_8{0%{opacity:0}22.70%{opacity:0}24.17%{opacity:1}31.86%{opacity:1}100%{opacity:0}}.v1_8{animation-name:v1_8}@keyframes b1_9{0%{transform:scaleX(0)}20.54%{transform:scaleX(0)}23.97%{transform:scaleX(1)}31.86%{transform:scaleX(1)}100%{transform:scaleX(0)}}.b1_9{animation-name:b1_9;transform-box:fill-box;transform-origin:left center;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes v1_9{0%{opacity:0}22.99%{opacity:0}24.46%{opacity:1}31.86%{opacity:1}100%{opacity:0}}.v1_9{animation-name:v1_9}@keyframes b1_10{0%{transform:scaleX(0)}20.83%{transform:scaleX(0)}24.26%{transform:scaleX(1)}31.86%{transform:scaleX(1)}100%{transform:scaleX(0)}}.b1_10{animation-name:b1_10;transform-box:fill-box;transform-origin:left center;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes v1_10{0%{opacity:0}23.28%{opacity:0}24.75%{opacity:1}31.86%{opacity:1}100%{opacity:0}}.v1_10{animation-name:v1_10}@keyframes b1_11{0%{transform:scaleX(0)}21.13%{transform:scaleX(0)}24.56%{transform:scaleX(1)}31.86%{transform:scaleX(1)}100%{transform:scaleX(0)}}.b1_11{animation-name:b1_11;transform-box:fill-box;transform-origin:left center;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes v1_11{0%{opacity:0}23.58%{opacity:0}25.05%{opacity:1}31.86%{opacity:1}100%{opacity:0}}.v1_11{animation-name:v1_11}@keyframes sc2{0%{opacity:0}33.33%{opacity:0}34.80%{opacity:1}48.53%{opacity:1}50.00%{opacity:0}100%{opacity:0}}.sc2{animation-name:sc2}@keyframes b2_0{0%{transform:scaleX(0)}34.56%{transform:scaleX(0)}37.99%{transform:scaleX(1)}48.53%{transform:scaleX(1)}100%{transform:scaleX(0)}}.b2_0{animation-name:b2_0;transform-box:fill-box;transform-origin:left center;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes v2_0{0%{opacity:0}37.01%{opacity:0}38.48%{opacity:1}48.53%{opacity:1}100%{opacity:0}}.v2_0{animation-name:v2_0}@keyframes b2_1{0%{transform:scaleX(0)}34.85%{transform:scaleX(0)}38.28%{transform:scaleX(1)}48.53%{transform:scaleX(1)}100%{transform:scaleX(0)}}.b2_1{animation-name:b2_1;transform-box:fill-box;transform-origin:left center;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes v2_1{0%{opacity:0}37.30%{opacity:0}38.77%{opacity:1}48.53%{opacity:1}100%{opacity:0}}.v2_1{animation-name:v2_1}@keyframes b2_2{0%{transform:scaleX(0)}35.15%{transform:scaleX(0)}38.58%{transform:scaleX(1)}48.53%{transform:scaleX(1)}100%{transform:scaleX(0)}}.b2_2{animation-name:b2_2;transform-box:fill-box;transform-origin:left center;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes v2_2{0%{opacity:0}37.60%{opacity:0}39.07%{opacity:1}48.53%{opacity:1}100%{opacity:0}}.v2_2{animation-name:v2_2}@keyframes b2_3{0%{transform:scaleX(0)}35.44%{transform:scaleX(0)}38.87%{transform:scaleX(1)}48.53%{transform:scaleX(1)}100%{transform:scaleX(0)}}.b2_3{animation-name:b2_3;transform-box:fill-box;transform-origin:left center;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes v2_3{0%{opacity:0}37.89%{opacity:0}39.36%{opacity:1}48.53%{opacity:1}100%{opacity:0}}.v2_3{animation-name:v2_3}@keyframes b2_4{0%{transform:scaleX(0)}35.74%{transform:scaleX(0)}39.17%{transform:scaleX(1)}48.53%{transform:scaleX(1)}100%{transform:scaleX(0)}}.b2_4{animation-name:b2_4;transform-box:fill-box;transform-origin:left center;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes v2_4{0%{opacity:0}38.19%{opacity:0}39.66%{opacity:1}48.53%{opacity:1}100%{opacity:0}}.v2_4{animation-name:v2_4}@keyframes b2_5{0%{transform:scaleX(0)}36.03%{transform:scaleX(0)}39.46%{transform:scaleX(1)}48.53%{transform:scaleX(1)}100%{transform:scaleX(0)}}.b2_5{animation-name:b2_5;transform-box:fill-box;transform-origin:left center;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes v2_5{0%{opacity:0}38.48%{opacity:0}39.95%{opacity:1}48.53%{opacity:1}100%{opacity:0}}.v2_5{animation-name:v2_5}@keyframes b2_6{0%{transform:scaleX(0)}36.32%{transform:scaleX(0)}39.75%{transform:scaleX(1)}48.53%{transform:scaleX(1)}100%{transform:scaleX(0)}}.b2_6{animation-name:b2_6;transform-box:fill-box;transform-origin:left center;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes v2_6{0%{opacity:0}38.77%{opacity:0}40.25%{opacity:1}48.53%{opacity:1}100%{opacity:0}}.v2_6{animation-name:v2_6}@keyframes b2_7{0%{transform:scaleX(0)}36.62%{transform:scaleX(0)}40.05%{transform:scaleX(1)}48.53%{transform:scaleX(1)}100%{transform:scaleX(0)}}.b2_7{animation-name:b2_7;transform-box:fill-box;transform-origin:left center;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes v2_7{0%{opacity:0}39.07%{opacity:0}40.54%{opacity:1}48.53%{opacity:1}100%{opacity:0}}.v2_7{animation-name:v2_7}@keyframes b2_8{0%{transform:scaleX(0)}36.91%{transform:scaleX(0)}40.34%{transform:scaleX(1)}48.53%{transform:scaleX(1)}100%{transform:scaleX(0)}}.b2_8{animation-name:b2_8;transform-box:fill-box;transform-origin:left center;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes v2_8{0%{opacity:0}39.36%{opacity:0}40.83%{opacity:1}48.53%{opacity:1}100%{opacity:0}}.v2_8{animation-name:v2_8}@keyframes b2_9{0%{transform:scaleX(0)}37.21%{transform:scaleX(0)}40.64%{transform:scaleX(1)}48.53%{transform:scaleX(1)}100%{transform:scaleX(0)}}.b2_9{animation-name:b2_9;transform-box:fill-box;transform-origin:left center;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes v2_9{0%{opacity:0}39.66%{opacity:0}41.13%{opacity:1}48.53%{opacity:1}100%{opacity:0}}.v2_9{animation-name:v2_9}@keyframes b2_10{0%{transform:scaleX(0)}37.50%{transform:scaleX(0)}40.93%{transform:scaleX(1)}48.53%{transform:scaleX(1)}100%{transform:scaleX(0)}}.b2_10{animation-name:b2_10;transform-box:fill-box;transform-origin:left center;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes v2_10{0%{opacity:0}39.95%{opacity:0}41.42%{opacity:1}48.53%{opacity:1}100%{opacity:0}}.v2_10{animation-name:v2_10}@keyframes b2_11{0%{transform:scaleX(0)}37.79%{transform:scaleX(0)}41.23%{transform:scaleX(1)}48.53%{transform:scaleX(1)}100%{transform:scaleX(0)}}.b2_11{animation-name:b2_11;transform-box:fill-box;transform-origin:left center;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes v2_11{0%{opacity:0}40.25%{opacity:0}41.72%{opacity:1}48.53%{opacity:1}100%{opacity:0}}.v2_11{animation-name:v2_11}@keyframes sc3{0%{opacity:0}50.00%{opacity:0}51.47%{opacity:1}65.20%{opacity:1}66.67%{opacity:0}100%{opacity:0}}.sc3{animation-name:sc3}@keyframes b3_0{0%{transform:scaleX(0)}51.23%{transform:scaleX(0)}54.66%{transform:scaleX(1)}65.20%{transform:scaleX(1)}100%{transform:scaleX(0)}}.b3_0{animation-name:b3_0;transform-box:fill-box;transform-origin:left center;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes v3_0{0%{opacity:0}53.68%{opacity:0}55.15%{opacity:1}65.20%{opacity:1}100%{opacity:0}}.v3_0{animation-name:v3_0}@keyframes b3_1{0%{transform:scaleX(0)}51.52%{transform:scaleX(0)}54.95%{transform:scaleX(1)}65.20%{transform:scaleX(1)}100%{transform:scaleX(0)}}.b3_1{animation-name:b3_1;transform-box:fill-box;transform-origin:left center;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes v3_1{0%{opacity:0}53.97%{opacity:0}55.44%{opacity:1}65.20%{opacity:1}100%{opacity:0}}.v3_1{animation-name:v3_1}@keyframes b3_2{0%{transform:scaleX(0)}51.81%{transform:scaleX(0)}55.25%{transform:scaleX(1)}65.20%{transform:scaleX(1)}100%{transform:scaleX(0)}}.b3_2{animation-name:b3_2;transform-box:fill-box;transform-origin:left center;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes v3_2{0%{opacity:0}54.26%{opacity:0}55.74%{opacity:1}65.20%{opacity:1}100%{opacity:0}}.v3_2{animation-name:v3_2}@keyframes b3_3{0%{transform:scaleX(0)}52.11%{transform:scaleX(0)}55.54%{transform:scaleX(1)}65.20%{transform:scaleX(1)}100%{transform:scaleX(0)}}.b3_3{animation-name:b3_3;transform-box:fill-box;transform-origin:left center;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes v3_3{0%{opacity:0}54.56%{opacity:0}56.03%{opacity:1}65.20%{opacity:1}100%{opacity:0}}.v3_3{animation-name:v3_3}@keyframes b3_4{0%{transform:scaleX(0)}52.40%{transform:scaleX(0)}55.83%{transform:scaleX(1)}65.20%{transform:scaleX(1)}100%{transform:scaleX(0)}}.b3_4{animation-name:b3_4;transform-box:fill-box;transform-origin:left center;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes v3_4{0%{opacity:0}54.85%{opacity:0}56.32%{opacity:1}65.20%{opacity:1}100%{opacity:0}}.v3_4{animation-name:v3_4}@keyframes b3_5{0%{transform:scaleX(0)}52.70%{transform:scaleX(0)}56.13%{transform:scaleX(1)}65.20%{transform:scaleX(1)}100%{transform:scaleX(0)}}.b3_5{animation-name:b3_5;transform-box:fill-box;transform-origin:left center;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes v3_5{0%{opacity:0}55.15%{opacity:0}56.62%{opacity:1}65.20%{opacity:1}100%{opacity:0}}.v3_5{animation-name:v3_5}@keyframes b3_6{0%{transform:scaleX(0)}52.99%{transform:scaleX(0)}56.42%{transform:scaleX(1)}65.20%{transform:scaleX(1)}100%{transform:scaleX(0)}}.b3_6{animation-name:b3_6;transform-box:fill-box;transform-origin:left center;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes v3_6{0%{opacity:0}55.44%{opacity:0}56.91%{opacity:1}65.20%{opacity:1}100%{opacity:0}}.v3_6{animation-name:v3_6}@keyframes b3_7{0%{transform:scaleX(0)}53.28%{transform:scaleX(0)}56.72%{transform:scaleX(1)}65.20%{transform:scaleX(1)}100%{transform:scaleX(0)}}.b3_7{animation-name:b3_7;transform-box:fill-box;transform-origin:left center;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes v3_7{0%{opacity:0}55.74%{opacity:0}57.21%{opacity:1}65.20%{opacity:1}100%{opacity:0}}.v3_7{animation-name:v3_7}@keyframes b3_8{0%{transform:scaleX(0)}53.58%{transform:scaleX(0)}57.01%{transform:scaleX(1)}65.20%{transform:scaleX(1)}100%{transform:scaleX(0)}}.b3_8{animation-name:b3_8;transform-box:fill-box;transform-origin:left center;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes v3_8{0%{opacity:0}56.03%{opacity:0}57.50%{opacity:1}65.20%{opacity:1}100%{opacity:0}}.v3_8{animation-name:v3_8}@keyframes b3_9{0%{transform:scaleX(0)}53.87%{transform:scaleX(0)}57.30%{transform:scaleX(1)}65.20%{transform:scaleX(1)}100%{transform:scaleX(0)}}.b3_9{animation-name:b3_9;transform-box:fill-box;transform-origin:left center;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes v3_9{0%{opacity:0}56.32%{opacity:0}57.79%{opacity:1}65.20%{opacity:1}100%{opacity:0}}.v3_9{animation-name:v3_9}@keyframes b3_10{0%{transform:scaleX(0)}54.17%{transform:scaleX(0)}57.60%{transform:scaleX(1)}65.20%{transform:scaleX(1)}100%{transform:scaleX(0)}}.b3_10{animation-name:b3_10;transform-box:fill-box;transform-origin:left center;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes v3_10{0%{opacity:0}56.62%{opacity:0}58.09%{opacity:1}65.20%{opacity:1}100%{opacity:0}}.v3_10{animation-name:v3_10}@keyframes b3_11{0%{transform:scaleX(0)}54.46%{transform:scaleX(0)}57.89%{transform:scaleX(1)}65.20%{transform:scaleX(1)}100%{transform:scaleX(0)}}.b3_11{animation-name:b3_11;transform-box:fill-box;transform-origin:left center;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes v3_11{0%{opacity:0}56.91%{opacity:0}58.38%{opacity:1}65.20%{opacity:1}100%{opacity:0}}.v3_11{animation-name:v3_11}@keyframes sc4{0%{opacity:0}66.67%{opacity:0}68.14%{opacity:1}81.86%{opacity:1}83.33%{opacity:0}100%{opacity:0}}.sc4{animation-name:sc4}@keyframes b4_0{0%{transform:scaleX(0)}67.89%{transform:scaleX(0)}71.32%{transform:scaleX(1)}81.86%{transform:scaleX(1)}100%{transform:scaleX(0)}}.b4_0{animation-name:b4_0;transform-box:fill-box;transform-origin:left center;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes v4_0{0%{opacity:0}70.34%{opacity:0}71.81%{opacity:1}81.86%{opacity:1}100%{opacity:0}}.v4_0{animation-name:v4_0}@keyframes b4_1{0%{transform:scaleX(0)}68.19%{transform:scaleX(0)}71.62%{transform:scaleX(1)}81.86%{transform:scaleX(1)}100%{transform:scaleX(0)}}.b4_1{animation-name:b4_1;transform-box:fill-box;transform-origin:left center;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes v4_1{0%{opacity:0}70.64%{opacity:0}72.11%{opacity:1}81.86%{opacity:1}100%{opacity:0}}.v4_1{animation-name:v4_1}@keyframes b4_2{0%{transform:scaleX(0)}68.48%{transform:scaleX(0)}71.91%{transform:scaleX(1)}81.86%{transform:scaleX(1)}100%{transform:scaleX(0)}}.b4_2{animation-name:b4_2;transform-box:fill-box;transform-origin:left center;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes v4_2{0%{opacity:0}70.93%{opacity:0}72.40%{opacity:1}81.86%{opacity:1}100%{opacity:0}}.v4_2{animation-name:v4_2}@keyframes b4_3{0%{transform:scaleX(0)}68.77%{transform:scaleX(0)}72.21%{transform:scaleX(1)}81.86%{transform:scaleX(1)}100%{transform:scaleX(0)}}.b4_3{animation-name:b4_3;transform-box:fill-box;transform-origin:left center;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes v4_3{0%{opacity:0}71.23%{opacity:0}72.70%{opacity:1}81.86%{opacity:1}100%{opacity:0}}.v4_3{animation-name:v4_3}@keyframes b4_4{0%{transform:scaleX(0)}69.07%{transform:scaleX(0)}72.50%{transform:scaleX(1)}81.86%{transform:scaleX(1)}100%{transform:scaleX(0)}}.b4_4{animation-name:b4_4;transform-box:fill-box;transform-origin:left center;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes v4_4{0%{opacity:0}71.52%{opacity:0}72.99%{opacity:1}81.86%{opacity:1}100%{opacity:0}}.v4_4{animation-name:v4_4}@keyframes b4_5{0%{transform:scaleX(0)}69.36%{transform:scaleX(0)}72.79%{transform:scaleX(1)}81.86%{transform:scaleX(1)}100%{transform:scaleX(0)}}.b4_5{animation-name:b4_5;transform-box:fill-box;transform-origin:left center;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes v4_5{0%{opacity:0}71.81%{opacity:0}73.28%{opacity:1}81.86%{opacity:1}100%{opacity:0}}.v4_5{animation-name:v4_5}@keyframes b4_6{0%{transform:scaleX(0)}69.66%{transform:scaleX(0)}73.09%{transform:scaleX(1)}81.86%{transform:scaleX(1)}100%{transform:scaleX(0)}}.b4_6{animation-name:b4_6;transform-box:fill-box;transform-origin:left center;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes v4_6{0%{opacity:0}72.11%{opacity:0}73.58%{opacity:1}81.86%{opacity:1}100%{opacity:0}}.v4_6{animation-name:v4_6}@keyframes b4_7{0%{transform:scaleX(0)}69.95%{transform:scaleX(0)}73.38%{transform:scaleX(1)}81.86%{transform:scaleX(1)}100%{transform:scaleX(0)}}.b4_7{animation-name:b4_7;transform-box:fill-box;transform-origin:left center;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes v4_7{0%{opacity:0}72.40%{opacity:0}73.87%{opacity:1}81.86%{opacity:1}100%{opacity:0}}.v4_7{animation-name:v4_7}@keyframes b4_8{0%{transform:scaleX(0)}70.25%{transform:scaleX(0)}73.68%{transform:scaleX(1)}81.86%{transform:scaleX(1)}100%{transform:scaleX(0)}}.b4_8{animation-name:b4_8;transform-box:fill-box;transform-origin:left center;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes v4_8{0%{opacity:0}72.70%{opacity:0}74.17%{opacity:1}81.86%{opacity:1}100%{opacity:0}}.v4_8{animation-name:v4_8}@keyframes b4_9{0%{transform:scaleX(0)}70.54%{transform:scaleX(0)}73.97%{transform:scaleX(1)}81.86%{transform:scaleX(1)}100%{transform:scaleX(0)}}.b4_9{animation-name:b4_9;transform-box:fill-box;transform-origin:left center;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes v4_9{0%{opacity:0}72.99%{opacity:0}74.46%{opacity:1}81.86%{opacity:1}100%{opacity:0}}.v4_9{animation-name:v4_9}@keyframes b4_10{0%{transform:scaleX(0)}70.83%{transform:scaleX(0)}74.26%{transform:scaleX(1)}81.86%{transform:scaleX(1)}100%{transform:scaleX(0)}}.b4_10{animation-name:b4_10;transform-box:fill-box;transform-origin:left center;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes v4_10{0%{opacity:0}73.28%{opacity:0}74.75%{opacity:1}81.86%{opacity:1}100%{opacity:0}}.v4_10{animation-name:v4_10}@keyframes b4_11{0%{transform:scaleX(0)}71.13%{transform:scaleX(0)}74.56%{transform:scaleX(1)}81.86%{transform:scaleX(1)}100%{transform:scaleX(0)}}.b4_11{animation-name:b4_11;transform-box:fill-box;transform-origin:left center;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes v4_11{0%{opacity:0}73.58%{opacity:0}75.05%{opacity:1}81.86%{opacity:1}100%{opacity:0}}.v4_11{animation-name:v4_11}@keyframes sc5{0%{opacity:0}83.33%{opacity:0}84.80%{opacity:1}98.53%{opacity:1}100.00%{opacity:0}100%{opacity:0}}.sc5{animation-name:sc5}@keyframes b5_0{0%{transform:scaleX(0)}84.56%{transform:scaleX(0)}87.99%{transform:scaleX(1)}98.53%{transform:scaleX(1)}100%{transform:scaleX(0)}}.b5_0{animation-name:b5_0;transform-box:fill-box;transform-origin:left center;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes v5_0{0%{opacity:0}87.01%{opacity:0}88.48%{opacity:1}98.53%{opacity:1}100%{opacity:0}}.v5_0{animation-name:v5_0}@keyframes b5_1{0%{transform:scaleX(0)}84.85%{transform:scaleX(0)}88.28%{transform:scaleX(1)}98.53%{transform:scaleX(1)}100%{transform:scaleX(0)}}.b5_1{animation-name:b5_1;transform-box:fill-box;transform-origin:left center;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes v5_1{0%{opacity:0}87.30%{opacity:0}88.77%{opacity:1}98.53%{opacity:1}100%{opacity:0}}.v5_1{animation-name:v5_1}@keyframes b5_2{0%{transform:scaleX(0)}85.15%{transform:scaleX(0)}88.58%{transform:scaleX(1)}98.53%{transform:scaleX(1)}100%{transform:scaleX(0)}}.b5_2{animation-name:b5_2;transform-box:fill-box;transform-origin:left center;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes v5_2{0%{opacity:0}87.60%{opacity:0}89.07%{opacity:1}98.53%{opacity:1}100%{opacity:0}}.v5_2{animation-name:v5_2}@keyframes b5_3{0%{transform:scaleX(0)}85.44%{transform:scaleX(0)}88.87%{transform:scaleX(1)}98.53%{transform:scaleX(1)}100%{transform:scaleX(0)}}.b5_3{animation-name:b5_3;transform-box:fill-box;transform-origin:left center;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes v5_3{0%{opacity:0}87.89%{opacity:0}89.36%{opacity:1}98.53%{opacity:1}100%{opacity:0}}.v5_3{animation-name:v5_3}@keyframes b5_4{0%{transform:scaleX(0)}85.74%{transform:scaleX(0)}89.17%{transform:scaleX(1)}98.53%{transform:scaleX(1)}100%{transform:scaleX(0)}}.b5_4{animation-name:b5_4;transform-box:fill-box;transform-origin:left center;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes v5_4{0%{opacity:0}88.19%{opacity:0}89.66%{opacity:1}98.53%{opacity:1}100%{opacity:0}}.v5_4{animation-name:v5_4}@keyframes b5_5{0%{transform:scaleX(0)}86.03%{transform:scaleX(0)}89.46%{transform:scaleX(1)}98.53%{transform:scaleX(1)}100%{transform:scaleX(0)}}.b5_5{animation-name:b5_5;transform-box:fill-box;transform-origin:left center;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes v5_5{0%{opacity:0}88.48%{opacity:0}89.95%{opacity:1}98.53%{opacity:1}100%{opacity:0}}.v5_5{animation-name:v5_5}@keyframes b5_6{0%{transform:scaleX(0)}86.32%{transform:scaleX(0)}89.75%{transform:scaleX(1)}98.53%{transform:scaleX(1)}100%{transform:scaleX(0)}}.b5_6{animation-name:b5_6;transform-box:fill-box;transform-origin:left center;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes v5_6{0%{opacity:0}88.77%{opacity:0}90.25%{opacity:1}98.53%{opacity:1}100%{opacity:0}}.v5_6{animation-name:v5_6}@keyframes b5_7{0%{transform:scaleX(0)}86.62%{transform:scaleX(0)}90.05%{transform:scaleX(1)}98.53%{transform:scaleX(1)}100%{transform:scaleX(0)}}.b5_7{animation-name:b5_7;transform-box:fill-box;transform-origin:left center;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes v5_7{0%{opacity:0}89.07%{opacity:0}90.54%{opacity:1}98.53%{opacity:1}100%{opacity:0}}.v5_7{animation-name:v5_7}@keyframes b5_8{0%{transform:scaleX(0)}86.91%{transform:scaleX(0)}90.34%{transform:scaleX(1)}98.53%{transform:scaleX(1)}100%{transform:scaleX(0)}}.b5_8{animation-name:b5_8;transform-box:fill-box;transform-origin:left center;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes v5_8{0%{opacity:0}89.36%{opacity:0}90.83%{opacity:1}98.53%{opacity:1}100%{opacity:0}}.v5_8{animation-name:v5_8}@keyframes b5_9{0%{transform:scaleX(0)}87.21%{transform:scaleX(0)}90.64%{transform:scaleX(1)}98.53%{transform:scaleX(1)}100%{transform:scaleX(0)}}.b5_9{animation-name:b5_9;transform-box:fill-box;transform-origin:left center;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes v5_9{0%{opacity:0}89.66%{opacity:0}91.13%{opacity:1}98.53%{opacity:1}100%{opacity:0}}.v5_9{animation-name:v5_9}@keyframes b5_10{0%{transform:scaleX(0)}87.50%{transform:scaleX(0)}90.93%{transform:scaleX(1)}98.53%{transform:scaleX(1)}100%{transform:scaleX(0)}}.b5_10{animation-name:b5_10;transform-box:fill-box;transform-origin:left center;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes v5_10{0%{opacity:0}89.95%{opacity:0}91.42%{opacity:1}98.53%{opacity:1}100%{opacity:0}}.v5_10{animation-name:v5_10}@keyframes b5_11{0%{transform:scaleX(0)}87.79%{transform:scaleX(0)}91.23%{transform:scaleX(1)}98.53%{transform:scaleX(1)}100%{transform:scaleX(0)}}.b5_11{animation-name:b5_11;transform-box:fill-box;transform-origin:left center;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes v5_11{0%{opacity:0}90.25%{opacity:0}91.72%{opacity:1}98.53%{opacity:1}100%{opacity:0}}.v5_11{animation-name:v5_11}@media (prefers-reduced-motion:reduce){.s{animation:none!important;opacity:1;transform:none}.rm{display:none}}</style>
+<rect width="900" height="482" rx="18" fill="#0e0e10"/>
+<text x="40" y="38" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="15" font-weight="600" fill="#e8e6df">Benchmarks</text>
+<text x="40" y="466" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10.5" fill="#85827a">Needle 3 through the shipped CQ2 binary, confidence gate on · baselines at f16 under vLLM · DeepSeek V4 Flash via its cloud API</text>
+<g class="s sc0">
+<text x="40" y="70" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="20" font-weight="600" fill="#e8e6df">Mobile Actions</text>
+<text x="40" y="88" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" fill="#85827a">exact call · 961 rows · accuracy (%)</text>
+<line x1="382.5" y1="102" x2="382.5" y2="418" stroke="#232326" stroke-width="1"/>
+<text x="382.5" y="436" text-anchor="middle" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">25</text>
+<line x1="515.0" y1="102" x2="515.0" y2="418" stroke="#232326" stroke-width="1"/>
+<text x="515.0" y="436" text-anchor="middle" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">50</text>
+<line x1="647.5" y1="102" x2="647.5" y2="418" stroke="#232326" stroke-width="1"/>
+<text x="647.5" y="436" text-anchor="middle" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">75</text>
+<line x1="780.0" y1="102" x2="780.0" y2="418" stroke="#232326" stroke-width="1"/>
+<text x="780.0" y="436" text-anchor="middle" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">100</text>
+<line x1="718.5" y1="96" x2="718.5" y2="422" stroke="#e8593d" stroke-width="1.2" stroke-dasharray="1 5" stroke-linecap="round" opacity="0.8"/>
+<text x="238" y="121" text-anchor="end" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="12.5" font-weight="400" fill="#e8e6df">DeepSeek V4 Flash</text>
+<text x="40" y="121" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">cloud</text>
+<rect class="s b0_0" x="250" y="109" width="468.5" height="17" rx="3" fill="#c9c6bd"/>
+<text class="s v0_0" x="726.5" y="122" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" font-weight="400" fill="#e8e6df">88.4</text>
+<text x="238" y="147" text-anchor="end" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="12.5" font-weight="600" fill="#e8e6df">Needle3-20L-121M</text>
+<text x="40" y="147" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">121M</text>
+<rect class="s b0_1" x="250" y="135" width="455.8" height="17" rx="3" fill="#e8593d"/>
+<text class="s v0_1" x="713.8" y="148" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" font-weight="600" fill="#e8e6df">86.0</text>
+<text x="238" y="173" text-anchor="end" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="12.5" font-weight="400" fill="#85827a">LFM2.5 1.2B</text>
+<text x="40" y="173" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">1.2B</text>
+<rect class="s b0_2" x="250" y="161" width="436.7" height="17" rx="3" fill="#4a4844"/>
+<text class="s v0_2" x="694.7" y="174" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" font-weight="400" fill="#85827a">82.4</text>
+<text x="238" y="199" text-anchor="end" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="12.5" font-weight="600" fill="#e8e6df">Needle3-16L-98M</text>
+<text x="40" y="199" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">98M</text>
+<rect class="s b0_3" x="250" y="187" width="427.7" height="17" rx="3" fill="#e8593d"/>
+<text class="s v0_3" x="685.7" y="200" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" font-weight="600" fill="#e8e6df">80.7</text>
+<text x="238" y="225" text-anchor="end" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="12.5" font-weight="400" fill="#85827a">Qwen3.5 0.8B</text>
+<text x="40" y="225" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">800M</text>
+<rect class="s b0_4" x="250" y="213" width="402.8" height="17" rx="3" fill="#4a4844"/>
+<text class="s v0_4" x="660.8" y="226" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" font-weight="400" fill="#85827a">76.0</text>
+<text x="238" y="251" text-anchor="end" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="12.5" font-weight="400" fill="#85827a">LFM2.5 350M</text>
+<text x="40" y="251" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">350M</text>
+<rect class="s b0_5" x="250" y="239" width="385.8" height="17" rx="3" fill="#4a4844"/>
+<text class="s v0_5" x="643.8" y="252" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" font-weight="400" fill="#85827a">72.8</text>
+<text x="238" y="277" text-anchor="end" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="12.5" font-weight="400" fill="#85827a">LFM2.5 230M</text>
+<text x="40" y="277" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">230M</text>
+<rect class="s b0_6" x="250" y="265" width="367.3" height="17" rx="3" fill="#4a4844"/>
+<text class="s v0_6" x="625.3" y="278" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" font-weight="400" fill="#85827a">69.3</text>
+<text x="238" y="303" text-anchor="end" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="12.5" font-weight="400" fill="#85827a">FunctionGemma 270M</text>
+<text x="40" y="303" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">270M</text>
+<rect class="s b0_7" x="250" y="291" width="345.0" height="17" rx="3" fill="#4a4844"/>
+<text class="s v0_7" x="603.0" y="304" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" font-weight="400" fill="#85827a">65.1</text>
+<text x="238" y="329" text-anchor="end" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="12.5" font-weight="400" fill="#85827a">Needle 2</text>
+<text x="40" y="329" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">45M</text>
+<rect class="s b0_8" x="250" y="317" width="336.6" height="17" rx="3" fill="#4a4844"/>
+<text class="s v0_8" x="594.5" y="330" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" font-weight="400" fill="#85827a">63.5</text>
+<text x="238" y="355" text-anchor="end" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="12.5" font-weight="400" fill="#85827a">Apple FM</text>
+<text x="40" y="355" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">3.0B</text>
+<rect class="s b0_9" x="250" y="343" width="305.3" height="17" rx="3" fill="#4a4844"/>
+<text class="s v0_9" x="563.3" y="356" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" font-weight="400" fill="#85827a">57.6</text>
+<text x="238" y="381" text-anchor="end" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="12.5" font-weight="600" fill="#e8e6df">Needle3-8L-52M</text>
+<text x="40" y="381" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">52M</text>
+<rect class="s b0_10" x="250" y="369" width="195.0" height="17" rx="3" fill="#e8593d"/>
+<text class="s v0_10" x="453.0" y="382" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" font-weight="600" fill="#e8e6df">36.8</text>
+<text x="238" y="407" text-anchor="end" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="12.5" font-weight="600" fill="#e8e6df">Needle3-4L-29M</text>
+<text x="40" y="407" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">29M</text>
+<rect class="s b0_11" x="250" y="395" width="62.0" height="17" rx="3" fill="#e8593d"/>
+<text class="s v0_11" x="320.0" y="408" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" font-weight="600" fill="#e8e6df">11.7</text>
+</g>
+<g class="s sc1">
+<text x="40" y="70" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="20" font-weight="600" fill="#e8e6df">DroidCall</text>
+<text x="40" y="88" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" fill="#85827a">exact calls in order · 200 rows · accuracy (%)</text>
+<line x1="382.5" y1="102" x2="382.5" y2="392" stroke="#232326" stroke-width="1"/>
+<text x="382.5" y="410" text-anchor="middle" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">25</text>
+<line x1="515.0" y1="102" x2="515.0" y2="392" stroke="#232326" stroke-width="1"/>
+<text x="515.0" y="410" text-anchor="middle" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">50</text>
+<line x1="647.5" y1="102" x2="647.5" y2="392" stroke="#232326" stroke-width="1"/>
+<text x="647.5" y="410" text-anchor="middle" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">75</text>
+<line x1="780.0" y1="102" x2="780.0" y2="392" stroke="#232326" stroke-width="1"/>
+<text x="780.0" y="410" text-anchor="middle" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">100</text>
+<line x1="570.6" y1="96" x2="570.6" y2="396" stroke="#e8593d" stroke-width="1.2" stroke-dasharray="1 5" stroke-linecap="round" opacity="0.8"/>
+<text x="238" y="121" text-anchor="end" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="12.5" font-weight="400" fill="#e8e6df">DeepSeek V4 Flash</text>
+<text x="40" y="121" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">cloud</text>
+<rect class="s b1_0" x="250" y="109" width="320.6" height="17" rx="3" fill="#c9c6bd"/>
+<text class="s v1_0" x="578.6" y="122" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" font-weight="400" fill="#e8e6df">60.5</text>
+<text x="238" y="147" text-anchor="end" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="12.5" font-weight="600" fill="#e8e6df">Needle3-20L-121M</text>
+<text x="40" y="147" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">121M</text>
+<rect class="s b1_1" x="250" y="135" width="249.1" height="17" rx="3" fill="#e8593d"/>
+<text class="s v1_1" x="507.1" y="148" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" font-weight="600" fill="#e8e6df">47.0</text>
+<text x="238" y="173" text-anchor="end" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="12.5" font-weight="600" fill="#e8e6df">Needle3-16L-98M</text>
+<text x="40" y="173" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">98M</text>
+<rect class="s b1_2" x="250" y="161" width="212.0" height="17" rx="3" fill="#e8593d"/>
+<text class="s v1_2" x="470.0" y="174" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" font-weight="600" fill="#e8e6df">40.0</text>
+<text x="238" y="199" text-anchor="end" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="12.5" font-weight="600" fill="#e8e6df">Needle3-8L-52M</text>
+<text x="40" y="199" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">52M</text>
+<rect class="s b1_3" x="250" y="187" width="193.4" height="17" rx="3" fill="#e8593d"/>
+<text class="s v1_3" x="451.4" y="200" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" font-weight="600" fill="#e8e6df">36.5</text>
+<text x="238" y="225" text-anchor="end" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="12.5" font-weight="400" fill="#85827a">LFM2.5 1.2B</text>
+<text x="40" y="225" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">1.2B</text>
+<rect class="s b1_4" x="250" y="213" width="188.2" height="17" rx="3" fill="#4a4844"/>
+<text class="s v1_4" x="446.1" y="226" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" font-weight="400" fill="#85827a">35.5</text>
+<text x="238" y="251" text-anchor="end" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="12.5" font-weight="400" fill="#85827a">LFM2.5 350M</text>
+<text x="40" y="251" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">350M</text>
+<rect class="s b1_5" x="250" y="239" width="172.2" height="17" rx="3" fill="#4a4844"/>
+<text class="s v1_5" x="430.2" y="252" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" font-weight="400" fill="#85827a">32.5</text>
+<text x="238" y="277" text-anchor="end" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="12.5" font-weight="400" fill="#85827a">Qwen3.5 0.8B</text>
+<text x="40" y="277" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">800M</text>
+<rect class="s b1_6" x="250" y="265" width="148.4" height="17" rx="3" fill="#4a4844"/>
+<text class="s v1_6" x="406.4" y="278" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" font-weight="400" fill="#85827a">28.0</text>
+<text x="238" y="303" text-anchor="end" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="12.5" font-weight="600" fill="#e8e6df">Needle3-4L-29M</text>
+<text x="40" y="303" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">29M</text>
+<rect class="s b1_7" x="250" y="291" width="111.3" height="17" rx="3" fill="#e8593d"/>
+<text class="s v1_7" x="369.3" y="304" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" font-weight="600" fill="#e8e6df">21.0</text>
+<text x="238" y="329" text-anchor="end" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="12.5" font-weight="400" fill="#85827a">Needle 2</text>
+<text x="40" y="329" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">45M</text>
+<rect class="s b1_8" x="250" y="317" width="90.1" height="17" rx="3" fill="#4a4844"/>
+<text class="s v1_8" x="348.1" y="330" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" font-weight="400" fill="#85827a">17.0</text>
+<text x="238" y="355" text-anchor="end" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="12.5" font-weight="400" fill="#85827a">FunctionGemma 270M</text>
+<text x="40" y="355" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">270M</text>
+<rect class="s b1_9" x="250" y="343" width="87.5" height="17" rx="3" fill="#4a4844"/>
+<text class="s v1_9" x="345.4" y="356" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" font-weight="400" fill="#85827a">16.5</text>
+<text x="238" y="381" text-anchor="end" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="12.5" font-weight="400" fill="#85827a">LFM2.5 230M</text>
+<text x="40" y="381" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">230M</text>
+<rect class="s b1_10" x="250" y="369" width="61.0" height="17" rx="3" fill="#4a4844"/>
+<text class="s v1_10" x="318.9" y="382" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" font-weight="400" fill="#85827a">11.5</text>
+</g>
+<g class="s sc2">
+<text x="40" y="70" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="20" font-weight="600" fill="#e8e6df">BFCL v4</text>
+<text x="40" y="88" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" fill="#85827a">AST match · 3,641 rows · accuracy (%)</text>
+<line x1="382.5" y1="102" x2="382.5" y2="366" stroke="#232326" stroke-width="1"/>
+<text x="382.5" y="384" text-anchor="middle" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">25</text>
+<line x1="515.0" y1="102" x2="515.0" y2="366" stroke="#232326" stroke-width="1"/>
+<text x="515.0" y="384" text-anchor="middle" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">50</text>
+<line x1="647.5" y1="102" x2="647.5" y2="366" stroke="#232326" stroke-width="1"/>
+<text x="647.5" y="384" text-anchor="middle" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">75</text>
+<line x1="780.0" y1="102" x2="780.0" y2="366" stroke="#232326" stroke-width="1"/>
+<text x="780.0" y="384" text-anchor="middle" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">100</text>
+<line x1="659.2" y1="96" x2="659.2" y2="370" stroke="#e8593d" stroke-width="1.2" stroke-dasharray="1 5" stroke-linecap="round" opacity="0.8"/>
+<text x="238" y="121" text-anchor="end" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="12.5" font-weight="400" fill="#e8e6df">DeepSeek V4 Flash</text>
+<text x="40" y="121" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">cloud</text>
+<rect class="s b2_0" x="250" y="109" width="409.2" height="17" rx="3" fill="#c9c6bd"/>
+<text class="s v2_0" x="667.2" y="122" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" font-weight="400" fill="#e8e6df">77.2</text>
+<text x="238" y="147" text-anchor="end" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="12.5" font-weight="400" fill="#85827a">LFM2.5 1.2B</text>
+<text x="40" y="147" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">1.2B</text>
+<rect class="s b2_1" x="250" y="135" width="328.6" height="17" rx="3" fill="#4a4844"/>
+<text class="s v2_1" x="586.6" y="148" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" font-weight="400" fill="#85827a">62.0</text>
+<text x="238" y="173" text-anchor="end" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="12.5" font-weight="400" fill="#85827a">LFM2.5 350M</text>
+<text x="40" y="173" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">350M</text>
+<rect class="s b2_2" x="250" y="161" width="313.2" height="17" rx="3" fill="#4a4844"/>
+<text class="s v2_2" x="571.2" y="174" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" font-weight="400" fill="#85827a">59.1</text>
+<text x="238" y="199" text-anchor="end" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="12.5" font-weight="400" fill="#85827a">Qwen3.5 0.8B</text>
+<text x="40" y="199" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">800M</text>
+<rect class="s b2_3" x="250" y="187" width="301.0" height="17" rx="3" fill="#4a4844"/>
+<text class="s v2_3" x="559.0" y="200" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" font-weight="400" fill="#85827a">56.8</text>
+<text x="238" y="225" text-anchor="end" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="12.5" font-weight="600" fill="#e8e6df">Needle3-20L-121M</text>
+<text x="40" y="225" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">121M</text>
+<rect class="s b2_4" x="250" y="213" width="266.1" height="17" rx="3" fill="#e8593d"/>
+<text class="s v2_4" x="524.1" y="226" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" font-weight="600" fill="#e8e6df">50.2</text>
+<text x="238" y="251" text-anchor="end" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="12.5" font-weight="400" fill="#85827a">FunctionGemma 270M</text>
+<text x="40" y="251" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">270M</text>
+<rect class="s b2_5" x="250" y="239" width="247.0" height="17" rx="3" fill="#4a4844"/>
+<text class="s v2_5" x="505.0" y="252" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" font-weight="400" fill="#85827a">46.6</text>
+<text x="238" y="277" text-anchor="end" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="12.5" font-weight="400" fill="#85827a">LFM2.5 230M</text>
+<text x="40" y="277" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">230M</text>
+<rect class="s b2_6" x="250" y="265" width="245.4" height="17" rx="3" fill="#4a4844"/>
+<text class="s v2_6" x="503.4" y="278" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" font-weight="400" fill="#85827a">46.3</text>
+<text x="238" y="303" text-anchor="end" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="12.5" font-weight="600" fill="#e8e6df">Needle3-16L-98M</text>
+<text x="40" y="303" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">98M</text>
+<rect class="s b2_7" x="250" y="291" width="218.9" height="17" rx="3" fill="#e8593d"/>
+<text class="s v2_7" x="476.9" y="304" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" font-weight="600" fill="#e8e6df">41.3</text>
+<text x="238" y="329" text-anchor="end" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="12.5" font-weight="600" fill="#e8e6df">Needle3-8L-52M</text>
+<text x="40" y="329" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">52M</text>
+<rect class="s b2_8" x="250" y="317" width="149.5" height="17" rx="3" fill="#e8593d"/>
+<text class="s v2_8" x="407.5" y="330" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" font-weight="600" fill="#e8e6df">28.2</text>
+<text x="238" y="355" text-anchor="end" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="12.5" font-weight="600" fill="#e8e6df">Needle3-4L-29M</text>
+<text x="40" y="355" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">29M</text>
+<rect class="s b2_9" x="250" y="343" width="103.3" height="17" rx="3" fill="#e8593d"/>
+<text class="s v2_9" x="361.4" y="356" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" font-weight="600" fill="#e8e6df">19.5</text>
+</g>
+<g class="s sc3">
+<text x="40" y="70" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="20" font-weight="600" fill="#e8e6df">DSTC8</text>
+<text x="40" y="88" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" fill="#85827a">field F1 · 1,813 turns · F1 (%)</text>
+<line x1="382.5" y1="102" x2="382.5" y2="366" stroke="#232326" stroke-width="1"/>
+<text x="382.5" y="384" text-anchor="middle" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">25</text>
+<line x1="515.0" y1="102" x2="515.0" y2="366" stroke="#232326" stroke-width="1"/>
+<text x="515.0" y="384" text-anchor="middle" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">50</text>
+<line x1="647.5" y1="102" x2="647.5" y2="366" stroke="#232326" stroke-width="1"/>
+<text x="647.5" y="384" text-anchor="middle" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">75</text>
+<line x1="780.0" y1="102" x2="780.0" y2="366" stroke="#232326" stroke-width="1"/>
+<text x="780.0" y="384" text-anchor="middle" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">100</text>
+<line x1="674.0" y1="96" x2="674.0" y2="370" stroke="#e8593d" stroke-width="1.2" stroke-dasharray="1 5" stroke-linecap="round" opacity="0.8"/>
+<text x="238" y="121" text-anchor="end" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="12.5" font-weight="400" fill="#e8e6df">DeepSeek V4 Flash</text>
+<text x="40" y="121" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">cloud</text>
+<rect class="s b3_0" x="250" y="109" width="424.0" height="17" rx="3" fill="#c9c6bd"/>
+<text class="s v3_0" x="682.0" y="122" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" font-weight="400" fill="#e8e6df">80.0</text>
+<text x="238" y="147" text-anchor="end" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="12.5" font-weight="400" fill="#85827a">LFM2.5 230M</text>
+<text x="40" y="147" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">230M</text>
+<rect class="s b3_1" x="250" y="135" width="280.9" height="17" rx="3" fill="#4a4844"/>
+<text class="s v3_1" x="538.9" y="148" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" font-weight="400" fill="#85827a">53.0</text>
+<text x="238" y="173" text-anchor="end" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="12.5" font-weight="400" fill="#85827a">Qwen3.5 0.8B</text>
+<text x="40" y="173" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">800M</text>
+<rect class="s b3_2" x="250" y="161" width="259.7" height="17" rx="3" fill="#4a4844"/>
+<text class="s v3_2" x="517.7" y="174" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" font-weight="400" fill="#85827a">49.0</text>
+<text x="238" y="199" text-anchor="end" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="12.5" font-weight="400" fill="#85827a">LFM2.5 1.2B</text>
+<text x="40" y="199" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">1.2B</text>
+<rect class="s b3_3" x="250" y="187" width="254.4" height="17" rx="3" fill="#4a4844"/>
+<text class="s v3_3" x="512.4" y="200" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" font-weight="400" fill="#85827a">48.0</text>
+<text x="238" y="225" text-anchor="end" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="12.5" font-weight="600" fill="#e8e6df">Needle3-20L-121M</text>
+<text x="40" y="225" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">121M</text>
+<rect class="s b3_4" x="250" y="213" width="215.7" height="17" rx="3" fill="#e8593d"/>
+<text class="s v3_4" x="473.7" y="226" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" font-weight="600" fill="#e8e6df">40.7</text>
+<text x="238" y="251" text-anchor="end" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="12.5" font-weight="600" fill="#e8e6df">Needle3-16L-98M</text>
+<text x="40" y="251" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">98M</text>
+<rect class="s b3_5" x="250" y="239" width="151.1" height="17" rx="3" fill="#e8593d"/>
+<text class="s v3_5" x="409.1" y="252" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" font-weight="600" fill="#e8e6df">28.5</text>
+<text x="238" y="277" text-anchor="end" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="12.5" font-weight="400" fill="#85827a">FunctionGemma 270M</text>
+<text x="40" y="277" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">270M</text>
+<rect class="s b3_6" x="250" y="265" width="143.1" height="17" rx="3" fill="#4a4844"/>
+<text class="s v3_6" x="401.1" y="278" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" font-weight="400" fill="#85827a">27.0</text>
+<text x="238" y="303" text-anchor="end" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="12.5" font-weight="400" fill="#85827a">LFM2.5 350M</text>
+<text x="40" y="303" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">350M</text>
+<rect class="s b3_7" x="250" y="291" width="106.0" height="17" rx="3" fill="#4a4844"/>
+<text class="s v3_7" x="364.0" y="304" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" font-weight="400" fill="#85827a">20.0</text>
+<text x="238" y="329" text-anchor="end" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="12.5" font-weight="600" fill="#e8e6df">Needle3-8L-52M</text>
+<text x="40" y="329" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">52M</text>
+<rect class="s b3_8" x="250" y="317" width="81.1" height="17" rx="3" fill="#e8593d"/>
+<text class="s v3_8" x="339.1" y="330" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" font-weight="600" fill="#e8e6df">15.3</text>
+<text x="238" y="355" text-anchor="end" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="12.5" font-weight="600" fill="#e8e6df">Needle3-4L-29M</text>
+<text x="40" y="355" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">29M</text>
+<rect class="s b3_9" x="250" y="343" width="36.6" height="17" rx="3" fill="#e8593d"/>
+<text class="s v3_9" x="294.6" y="356" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" font-weight="600" fill="#e8e6df">6.9</text>
+</g>
+<g class="s sc4">
+<text x="40" y="70" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="20" font-weight="600" fill="#e8e6df">SNIPS gold</text>
+<text x="40" y="88" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" fill="#85827a">field F1 · 700 rows · F1 (%)</text>
+<line x1="382.5" y1="102" x2="382.5" y2="366" stroke="#232326" stroke-width="1"/>
+<text x="382.5" y="384" text-anchor="middle" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">25</text>
+<line x1="515.0" y1="102" x2="515.0" y2="366" stroke="#232326" stroke-width="1"/>
+<text x="515.0" y="384" text-anchor="middle" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">50</text>
+<line x1="647.5" y1="102" x2="647.5" y2="366" stroke="#232326" stroke-width="1"/>
+<text x="647.5" y="384" text-anchor="middle" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">75</text>
+<line x1="780.0" y1="102" x2="780.0" y2="366" stroke="#232326" stroke-width="1"/>
+<text x="780.0" y="384" text-anchor="middle" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">100</text>
+<line x1="617.8" y1="96" x2="617.8" y2="370" stroke="#e8593d" stroke-width="1.2" stroke-dasharray="1 5" stroke-linecap="round" opacity="0.8"/>
+<text x="238" y="121" text-anchor="end" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="12.5" font-weight="400" fill="#e8e6df">DeepSeek V4 Flash</text>
+<text x="40" y="121" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">cloud</text>
+<rect class="s b4_0" x="250" y="109" width="367.8" height="17" rx="3" fill="#c9c6bd"/>
+<text class="s v4_0" x="625.8" y="122" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" font-weight="400" fill="#e8e6df">69.4</text>
+<text x="238" y="147" text-anchor="end" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="12.5" font-weight="400" fill="#85827a">LFM2.5 1.2B</text>
+<text x="40" y="147" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">1.2B</text>
+<rect class="s b4_1" x="250" y="135" width="227.9" height="17" rx="3" fill="#4a4844"/>
+<text class="s v4_1" x="485.9" y="148" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" font-weight="400" fill="#85827a">43.0</text>
+<text x="238" y="173" text-anchor="end" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="12.5" font-weight="400" fill="#85827a">Qwen3.5 0.8B</text>
+<text x="40" y="173" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">800M</text>
+<rect class="s b4_2" x="250" y="161" width="185.5" height="17" rx="3" fill="#4a4844"/>
+<text class="s v4_2" x="443.5" y="174" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" font-weight="400" fill="#85827a">35.0</text>
+<text x="238" y="199" text-anchor="end" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="12.5" font-weight="400" fill="#85827a">LFM2.5 350M</text>
+<text x="40" y="199" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">350M</text>
+<rect class="s b4_3" x="250" y="187" width="180.2" height="17" rx="3" fill="#4a4844"/>
+<text class="s v4_3" x="438.2" y="200" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" font-weight="400" fill="#85827a">34.0</text>
+<text x="238" y="225" text-anchor="end" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="12.5" font-weight="600" fill="#e8e6df">Needle3-20L-121M</text>
+<text x="40" y="225" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">121M</text>
+<rect class="s b4_4" x="250" y="213" width="160.1" height="17" rx="3" fill="#e8593d"/>
+<text class="s v4_4" x="418.1" y="226" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" font-weight="600" fill="#e8e6df">30.2</text>
+<text x="238" y="251" text-anchor="end" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="12.5" font-weight="400" fill="#85827a">FunctionGemma 270M</text>
+<text x="40" y="251" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">270M</text>
+<rect class="s b4_5" x="250" y="239" width="153.7" height="17" rx="3" fill="#4a4844"/>
+<text class="s v4_5" x="411.7" y="252" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" font-weight="400" fill="#85827a">29.0</text>
+<text x="238" y="277" text-anchor="end" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="12.5" font-weight="400" fill="#85827a">LFM2.5 230M</text>
+<text x="40" y="277" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">230M</text>
+<rect class="s b4_6" x="250" y="265" width="143.1" height="17" rx="3" fill="#4a4844"/>
+<text class="s v4_6" x="401.1" y="278" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" font-weight="400" fill="#85827a">27.0</text>
+<text x="238" y="303" text-anchor="end" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="12.5" font-weight="600" fill="#e8e6df">Needle3-16L-98M</text>
+<text x="40" y="303" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">98M</text>
+<rect class="s b4_7" x="250" y="291" width="124.5" height="17" rx="3" fill="#e8593d"/>
+<text class="s v4_7" x="382.6" y="304" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" font-weight="600" fill="#e8e6df">23.5</text>
+<text x="238" y="329" text-anchor="end" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="12.5" font-weight="600" fill="#e8e6df">Needle3-8L-52M</text>
+<text x="40" y="329" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">52M</text>
+<rect class="s b4_8" x="250" y="317" width="88.0" height="17" rx="3" fill="#e8593d"/>
+<text class="s v4_8" x="346.0" y="330" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" font-weight="600" fill="#e8e6df">16.6</text>
+<text x="238" y="355" text-anchor="end" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="12.5" font-weight="600" fill="#e8e6df">Needle3-4L-29M</text>
+<text x="40" y="355" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">29M</text>
+<rect class="s b4_9" x="250" y="343" width="40.8" height="17" rx="3" fill="#e8593d"/>
+<text class="s v4_9" x="298.8" y="356" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" font-weight="600" fill="#e8e6df">7.7</text>
+</g>
+<g class="s sc5">
+<text x="40" y="70" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="20" font-weight="600" fill="#e8e6df">SNIPS 7-way</text>
+<text x="40" y="88" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" fill="#85827a">field F1 · 700 rows · F1 (%)</text>
+<line x1="382.5" y1="102" x2="382.5" y2="366" stroke="#232326" stroke-width="1"/>
+<text x="382.5" y="384" text-anchor="middle" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">25</text>
+<line x1="515.0" y1="102" x2="515.0" y2="366" stroke="#232326" stroke-width="1"/>
+<text x="515.0" y="384" text-anchor="middle" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">50</text>
+<line x1="647.5" y1="102" x2="647.5" y2="366" stroke="#232326" stroke-width="1"/>
+<text x="647.5" y="384" text-anchor="middle" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">75</text>
+<line x1="780.0" y1="102" x2="780.0" y2="366" stroke="#232326" stroke-width="1"/>
+<text x="780.0" y="384" text-anchor="middle" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">100</text>
+<line x1="603.5" y1="96" x2="603.5" y2="370" stroke="#e8593d" stroke-width="1.2" stroke-dasharray="1 5" stroke-linecap="round" opacity="0.8"/>
+<text x="238" y="121" text-anchor="end" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="12.5" font-weight="400" fill="#e8e6df">DeepSeek V4 Flash</text>
+<text x="40" y="121" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">cloud</text>
+<rect class="s b5_0" x="250" y="109" width="353.5" height="17" rx="3" fill="#c9c6bd"/>
+<text class="s v5_0" x="611.5" y="122" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" font-weight="400" fill="#e8e6df">66.7</text>
+<text x="238" y="147" text-anchor="end" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="12.5" font-weight="400" fill="#85827a">LFM2.5 1.2B</text>
+<text x="40" y="147" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">1.2B</text>
+<rect class="s b5_1" x="250" y="135" width="201.4" height="17" rx="3" fill="#4a4844"/>
+<text class="s v5_1" x="459.4" y="148" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" font-weight="400" fill="#85827a">38.0</text>
+<text x="238" y="173" text-anchor="end" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="12.5" font-weight="400" fill="#85827a">Qwen3.5 0.8B</text>
+<text x="40" y="173" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">800M</text>
+<rect class="s b5_2" x="250" y="161" width="180.2" height="17" rx="3" fill="#4a4844"/>
+<text class="s v5_2" x="438.2" y="174" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" font-weight="400" fill="#85827a">34.0</text>
+<text x="238" y="199" text-anchor="end" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="12.5" font-weight="400" fill="#85827a">LFM2.5 350M</text>
+<text x="40" y="199" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">350M</text>
+<rect class="s b5_3" x="250" y="187" width="153.7" height="17" rx="3" fill="#4a4844"/>
+<text class="s v5_3" x="411.7" y="200" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" font-weight="400" fill="#85827a">29.0</text>
+<text x="238" y="225" text-anchor="end" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="12.5" font-weight="600" fill="#e8e6df">Needle3-20L-121M</text>
+<text x="40" y="225" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">121M</text>
+<rect class="s b5_4" x="250" y="213" width="130.9" height="17" rx="3" fill="#e8593d"/>
+<text class="s v5_4" x="388.9" y="226" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" font-weight="600" fill="#e8e6df">24.7</text>
+<text x="238" y="251" text-anchor="end" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="12.5" font-weight="400" fill="#85827a">LFM2.5 230M</text>
+<text x="40" y="251" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">230M</text>
+<rect class="s b5_5" x="250" y="239" width="116.6" height="17" rx="3" fill="#4a4844"/>
+<text class="s v5_5" x="374.6" y="252" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" font-weight="400" fill="#85827a">22.0</text>
+<text x="238" y="277" text-anchor="end" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="12.5" font-weight="600" fill="#e8e6df">Needle3-16L-98M</text>
+<text x="40" y="277" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">98M</text>
+<rect class="s b5_6" x="250" y="265" width="101.8" height="17" rx="3" fill="#e8593d"/>
+<text class="s v5_6" x="359.8" y="278" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" font-weight="600" fill="#e8e6df">19.2</text>
+<text x="238" y="303" text-anchor="end" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="12.5" font-weight="400" fill="#85827a">FunctionGemma 270M</text>
+<text x="40" y="303" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">270M</text>
+<rect class="s b5_7" x="250" y="291" width="74.2" height="17" rx="3" fill="#4a4844"/>
+<text class="s v5_7" x="332.2" y="304" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" font-weight="400" fill="#85827a">14.0</text>
+<text x="238" y="329" text-anchor="end" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="12.5" font-weight="600" fill="#e8e6df">Needle3-8L-52M</text>
+<text x="40" y="329" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">52M</text>
+<rect class="s b5_8" x="250" y="317" width="53.5" height="17" rx="3" fill="#e8593d"/>
+<text class="s v5_8" x="311.5" y="330" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" font-weight="600" fill="#e8e6df">10.1</text>
+<text x="238" y="355" text-anchor="end" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="12.5" font-weight="600" fill="#e8e6df">Needle3-4L-29M</text>
+<text x="40" y="355" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10" fill="#4a4844">29M</text>
+<rect class="s b5_9" x="250" y="343" width="22.8" height="17" rx="3" fill="#e8593d"/>
+<text class="s v5_9" x="280.8" y="356" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" font-weight="600" fill="#e8e6df">4.3</text>
+</g>
+</svg>
\ No newline at end of file
diff --git a/assets/deploy.svg b/assets/deploy.svg
new file mode 100644
index 0000000..01b6159
--- /dev/null
+++ b/assets/deploy.svg
@@ -0,0 +1,73 @@
+<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 900 529" width="900" height="529" role="img" aria-label="The thirteen platform folders Needle ships, each with an engine under 1 MB">
+<style>svg{--t0:0.0s}.s{animation-duration:14.3s;animation-iteration-count:infinite;animation-fill-mode:both;animation-delay:calc(0s - var(--t0))}@keyframes t0{0%{stroke:#232326;fill:#141416}0.00%{stroke:#232326;fill:#141416}1.05%{stroke:#e8593d;fill:#1d1512}6.99%{stroke:#e8593d;fill:#1d1512}7.69%{stroke:#232326;fill:#141416}100%{stroke:#232326;fill:#141416}}.t0{animation-name:t0}@keyframes f0{0%{fill:#85827a}0.00%{fill:#85827a}1.05%{fill:#e8593d}6.99%{fill:#e8593d}7.69%{fill:#85827a}100%{fill:#85827a}}.f0{animation-name:f0}@keyframes c0{0%{opacity:0}0.00%{opacity:0}0.84%{opacity:1}7.13%{opacity:1}7.69%{opacity:0}100%{opacity:0}}.c0{animation-name:c0}@keyframes t1{0%{stroke:#232326;fill:#141416}7.69%{stroke:#232326;fill:#141416}8.74%{stroke:#e8593d;fill:#1d1512}14.69%{stroke:#e8593d;fill:#1d1512}15.38%{stroke:#232326;fill:#141416}100%{stroke:#232326;fill:#141416}}.t1{animation-name:t1}@keyframes f1{0%{fill:#85827a}7.69%{fill:#85827a}8.74%{fill:#e8593d}14.69%{fill:#e8593d}15.38%{fill:#85827a}100%{fill:#85827a}}.f1{animation-name:f1}@keyframes c1{0%{opacity:0}7.69%{opacity:0}8.53%{opacity:1}14.83%{opacity:1}15.38%{opacity:0}100%{opacity:0}}.c1{animation-name:c1}@keyframes t2{0%{stroke:#232326;fill:#141416}15.38%{stroke:#232326;fill:#141416}16.43%{stroke:#e8593d;fill:#1d1512}22.38%{stroke:#e8593d;fill:#1d1512}23.08%{stroke:#232326;fill:#141416}100%{stroke:#232326;fill:#141416}}.t2{animation-name:t2}@keyframes f2{0%{fill:#85827a}15.38%{fill:#85827a}16.43%{fill:#e8593d}22.38%{fill:#e8593d}23.08%{fill:#85827a}100%{fill:#85827a}}.f2{animation-name:f2}@keyframes c2{0%{opacity:0}15.38%{opacity:0}16.22%{opacity:1}22.52%{opacity:1}23.08%{opacity:0}100%{opacity:0}}.c2{animation-name:c2}@keyframes t3{0%{stroke:#232326;fill:#141416}23.08%{stroke:#232326;fill:#141416}24.13%{stroke:#e8593d;fill:#1d1512}30.07%{stroke:#e8593d;fill:#1d1512}30.77%{stroke:#232326;fill:#141416}100%{stroke:#232326;fill:#141416}}.t3{animation-name:t3}@keyframes f3{0%{fill:#85827a}23.08%{fill:#85827a}24.13%{fill:#e8593d}30.07%{fill:#e8593d}30.77%{fill:#85827a}100%{fill:#85827a}}.f3{animation-name:f3}@keyframes c3{0%{opacity:0}23.08%{opacity:0}23.92%{opacity:1}30.21%{opacity:1}30.77%{opacity:0}100%{opacity:0}}.c3{animation-name:c3}@keyframes t4{0%{stroke:#232326;fill:#141416}30.77%{stroke:#232326;fill:#141416}31.82%{stroke:#e8593d;fill:#1d1512}37.76%{stroke:#e8593d;fill:#1d1512}38.46%{stroke:#232326;fill:#141416}100%{stroke:#232326;fill:#141416}}.t4{animation-name:t4}@keyframes f4{0%{fill:#85827a}30.77%{fill:#85827a}31.82%{fill:#e8593d}37.76%{fill:#e8593d}38.46%{fill:#85827a}100%{fill:#85827a}}.f4{animation-name:f4}@keyframes c4{0%{opacity:0}30.77%{opacity:0}31.61%{opacity:1}37.90%{opacity:1}38.46%{opacity:0}100%{opacity:0}}.c4{animation-name:c4}@keyframes t5{0%{stroke:#232326;fill:#141416}38.46%{stroke:#232326;fill:#141416}39.51%{stroke:#e8593d;fill:#1d1512}45.45%{stroke:#e8593d;fill:#1d1512}46.15%{stroke:#232326;fill:#141416}100%{stroke:#232326;fill:#141416}}.t5{animation-name:t5}@keyframes f5{0%{fill:#85827a}38.46%{fill:#85827a}39.51%{fill:#e8593d}45.45%{fill:#e8593d}46.15%{fill:#85827a}100%{fill:#85827a}}.f5{animation-name:f5}@keyframes c5{0%{opacity:0}38.46%{opacity:0}39.30%{opacity:1}45.59%{opacity:1}46.15%{opacity:0}100%{opacity:0}}.c5{animation-name:c5}@keyframes t6{0%{stroke:#232326;fill:#141416}46.15%{stroke:#232326;fill:#141416}47.20%{stroke:#e8593d;fill:#1d1512}53.15%{stroke:#e8593d;fill:#1d1512}53.85%{stroke:#232326;fill:#141416}100%{stroke:#232326;fill:#141416}}.t6{animation-name:t6}@keyframes f6{0%{fill:#85827a}46.15%{fill:#85827a}47.20%{fill:#e8593d}53.15%{fill:#e8593d}53.85%{fill:#85827a}100%{fill:#85827a}}.f6{animation-name:f6}@keyframes c6{0%{opacity:0}46.15%{opacity:0}46.99%{opacity:1}53.29%{opacity:1}53.85%{opacity:0}100%{opacity:0}}.c6{animation-name:c6}@keyframes t7{0%{stroke:#232326;fill:#141416}53.85%{stroke:#232326;fill:#141416}54.90%{stroke:#e8593d;fill:#1d1512}60.84%{stroke:#e8593d;fill:#1d1512}61.54%{stroke:#232326;fill:#141416}100%{stroke:#232326;fill:#141416}}.t7{animation-name:t7}@keyframes f7{0%{fill:#85827a}53.85%{fill:#85827a}54.90%{fill:#e8593d}60.84%{fill:#e8593d}61.54%{fill:#85827a}100%{fill:#85827a}}.f7{animation-name:f7}@keyframes c7{0%{opacity:0}53.85%{opacity:0}54.69%{opacity:1}60.98%{opacity:1}61.54%{opacity:0}100%{opacity:0}}.c7{animation-name:c7}@keyframes t8{0%{stroke:#232326;fill:#141416}61.54%{stroke:#232326;fill:#141416}62.59%{stroke:#e8593d;fill:#1d1512}68.53%{stroke:#e8593d;fill:#1d1512}69.23%{stroke:#232326;fill:#141416}100%{stroke:#232326;fill:#141416}}.t8{animation-name:t8}@keyframes f8{0%{fill:#85827a}61.54%{fill:#85827a}62.59%{fill:#e8593d}68.53%{fill:#e8593d}69.23%{fill:#85827a}100%{fill:#85827a}}.f8{animation-name:f8}@keyframes c8{0%{opacity:0}61.54%{opacity:0}62.38%{opacity:1}68.67%{opacity:1}69.23%{opacity:0}100%{opacity:0}}.c8{animation-name:c8}@keyframes t9{0%{stroke:#232326;fill:#141416}69.23%{stroke:#232326;fill:#141416}70.28%{stroke:#e8593d;fill:#1d1512}76.22%{stroke:#e8593d;fill:#1d1512}76.92%{stroke:#232326;fill:#141416}100%{stroke:#232326;fill:#141416}}.t9{animation-name:t9}@keyframes f9{0%{fill:#85827a}69.23%{fill:#85827a}70.28%{fill:#e8593d}76.22%{fill:#e8593d}76.92%{fill:#85827a}100%{fill:#85827a}}.f9{animation-name:f9}@keyframes c9{0%{opacity:0}69.23%{opacity:0}70.07%{opacity:1}76.36%{opacity:1}76.92%{opacity:0}100%{opacity:0}}.c9{animation-name:c9}@keyframes t10{0%{stroke:#232326;fill:#141416}76.92%{stroke:#232326;fill:#141416}77.97%{stroke:#e8593d;fill:#1d1512}83.92%{stroke:#e8593d;fill:#1d1512}84.62%{stroke:#232326;fill:#141416}100%{stroke:#232326;fill:#141416}}.t10{animation-name:t10}@keyframes f10{0%{fill:#85827a}76.92%{fill:#85827a}77.97%{fill:#e8593d}83.92%{fill:#e8593d}84.62%{fill:#85827a}100%{fill:#85827a}}.f10{animation-name:f10}@keyframes c10{0%{opacity:0}76.92%{opacity:0}77.76%{opacity:1}84.06%{opacity:1}84.62%{opacity:0}100%{opacity:0}}.c10{animation-name:c10}@keyframes t11{0%{stroke:#232326;fill:#141416}84.62%{stroke:#232326;fill:#141416}85.66%{stroke:#e8593d;fill:#1d1512}91.61%{stroke:#e8593d;fill:#1d1512}92.31%{stroke:#232326;fill:#141416}100%{stroke:#232326;fill:#141416}}.t11{animation-name:t11}@keyframes f11{0%{fill:#85827a}84.62%{fill:#85827a}85.66%{fill:#e8593d}91.61%{fill:#e8593d}92.31%{fill:#85827a}100%{fill:#85827a}}.f11{animation-name:f11}@keyframes c11{0%{opacity:0}84.62%{opacity:0}85.45%{opacity:1}91.75%{opacity:1}92.31%{opacity:0}100%{opacity:0}}.c11{animation-name:c11}@keyframes t12{0%{stroke:#232326;fill:#141416}92.31%{stroke:#232326;fill:#141416}93.36%{stroke:#e8593d;fill:#1d1512}99.30%{stroke:#e8593d;fill:#1d1512}100.00%{stroke:#232326;fill:#141416}100%{stroke:#232326;fill:#141416}}.t12{animation-name:t12}@keyframes f12{0%{fill:#85827a}92.31%{fill:#85827a}93.36%{fill:#e8593d}99.30%{fill:#e8593d}100.00%{fill:#85827a}100%{fill:#85827a}}.f12{animation-name:f12}@keyframes c12{0%{opacity:0}92.31%{opacity:0}93.15%{opacity:1}99.44%{opacity:1}100.00%{opacity:0}100%{opacity:0}}.c12{animation-name:c12}@media (prefers-reduced-motion:reduce){.s{animation:none!important}.c0{opacity:1}}</style>
+<rect width="900" height="529" rx="18" fill="#0e0e10"/>
+<text x="40" y="38" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="15" font-weight="600" fill="#e8e6df">One engine per platform folder</text>
+<text x="860" y="38" text-anchor="end" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" fill="#85827a">each under 1 MB, each loading the same needle3.cact</text>
+<rect class="s t0" x="40" y="72" width="266" height="62" rx="9" fill="#141416" stroke="#232326" stroke-width="1.2"/>
+<text x="54" y="94" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="12.5" fill="#e8e6df">Mac (Apple Silicon)</text>
+<text class="s f0" x="54" y="111" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10.5" fill="#85827a">macos-arm64</text>
+<text x="54" y="125" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="9.5" fill="#4a4844">needle · libneedle.a · needle.h</text>
+<rect class="s t1" x="317" y="72" width="266" height="62" rx="9" fill="#141416" stroke="#232326" stroke-width="1.2"/>
+<text x="331" y="94" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="12.5" fill="#e8e6df">Linux x86-64 (PC, server)</text>
+<text class="s f1" x="331" y="111" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10.5" fill="#85827a">linux-x86_64</text>
+<text x="331" y="125" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="9.5" fill="#4a4844">needle · libneedle.a · needle.h</text>
+<rect class="s t2" x="594" y="72" width="266" height="62" rx="9" fill="#141416" stroke="#232326" stroke-width="1.2"/>
+<text x="608" y="94" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="12.5" fill="#e8e6df">Linux ARM64 (Raspberry Pi)</text>
+<text class="s f2" x="608" y="111" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10.5" fill="#85827a">linux-arm64</text>
+<text x="608" y="125" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="9.5" fill="#4a4844">needle · libneedle.a · needle.h</text>
+<rect class="s t3" x="40" y="145" width="266" height="62" rx="9" fill="#141416" stroke="#232326" stroke-width="1.2"/>
+<text x="54" y="167" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="12.5" fill="#e8e6df">Linux ARMv7 (32-bit)</text>
+<text class="s f3" x="54" y="184" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10.5" fill="#85827a">linux-armv7</text>
+<text x="54" y="198" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="9.5" fill="#4a4844">needle · libneedle.a · needle.h</text>
+<rect class="s t4" x="317" y="145" width="266" height="62" rx="9" fill="#141416" stroke="#232326" stroke-width="1.2"/>
+<text x="331" y="167" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="12.5" fill="#e8e6df">Linux RISC-V</text>
+<text class="s f4" x="331" y="184" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10.5" fill="#85827a">linux-riscv64</text>
+<text x="331" y="198" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="9.5" fill="#4a4844">needle · libneedle.a · needle.h</text>
+<rect class="s t5" x="594" y="145" width="266" height="62" rx="9" fill="#141416" stroke="#232326" stroke-width="1.2"/>
+<text x="608" y="167" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="12.5" fill="#e8e6df">Linux MIPS32el (cameras, routers)</text>
+<text class="s f5" x="608" y="184" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10.5" fill="#85827a">linux-mipsel</text>
+<text x="608" y="198" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="9.5" fill="#4a4844">needle · libneedle.a · needle.h</text>
+<rect class="s t6" x="40" y="218" width="266" height="62" rx="9" fill="#141416" stroke="#232326" stroke-width="1.2"/>
+<text x="54" y="240" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="12.5" fill="#e8e6df">Windows x64</text>
+<text class="s f6" x="54" y="257" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10.5" fill="#85827a">windows-x86_64</text>
+<text x="54" y="271" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="9.5" fill="#4a4844">needle.exe · libneedle.a · needle.h</text>
+<rect class="s t7" x="317" y="218" width="266" height="62" rx="9" fill="#141416" stroke="#232326" stroke-width="1.2"/>
+<text x="331" y="240" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="12.5" fill="#e8e6df">Windows ARM</text>
+<text class="s f7" x="331" y="257" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10.5" fill="#85827a">windows-arm64</text>
+<text x="331" y="271" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="9.5" fill="#4a4844">needle.exe · libneedle.a · needle.h</text>
+<rect class="s t8" x="594" y="218" width="266" height="62" rx="9" fill="#141416" stroke="#232326" stroke-width="1.2"/>
+<text x="608" y="240" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="12.5" fill="#e8e6df">Android</text>
+<text class="s f8" x="608" y="257" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10.5" fill="#85827a">android-arm64 · armv7 · riscv64</text>
+<text x="608" y="271" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="9.5" fill="#4a4844">needle · libneedle.a · needle.h</text>
+<rect class="s t9" x="40" y="291" width="266" height="62" rx="9" fill="#141416" stroke="#232326" stroke-width="1.2"/>
+<text x="54" y="313" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="12.5" fill="#e8e6df">iOS</text>
+<text class="s f9" x="54" y="330" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10.5" fill="#85827a">ios-arm64 · ios-sim-arm64</text>
+<text x="54" y="344" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="9.5" fill="#4a4844">libneedle.a · needle.h</text>
+<rect class="s t10" x="317" y="291" width="266" height="62" rx="9" fill="#141416" stroke="#232326" stroke-width="1.2"/>
+<text x="331" y="313" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="12.5" fill="#e8e6df">watchOS, tvOS</text>
+<text class="s f10" x="331" y="330" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10.5" fill="#85827a">watchos-arm64 · tvos-arm64</text>
+<text x="331" y="344" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="9.5" fill="#4a4844">libneedle.a · needle.h</text>
+<rect class="s t11" x="594" y="291" width="266" height="62" rx="9" fill="#141416" stroke="#232326" stroke-width="1.2"/>
+<text x="608" y="313" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="12.5" fill="#e8e6df">Browser, Node</text>
+<text class="s f11" x="608" y="330" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10.5" fill="#85827a">wasm</text>
+<text x="608" y="344" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="9.5" fill="#4a4844">needle.js · needle.wasm · needle.h</text>
+<rect class="s t12" x="40" y="364" width="266" height="62" rx="9" fill="#141416" stroke="#232326" stroke-width="1.2"/>
+<text x="54" y="386" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="12.5" fill="#e8e6df">WASI component hosts</text>
+<text class="s f12" x="54" y="403" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10.5" fill="#85827a">wasm-component</text>
+<text x="54" y="417" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="9.5" fill="#4a4844">needle.component.wasm · needle.wit</text>
+<text x="40" y="461" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="12" fill="#85827a">$</text>
+<text class="s c0" x="58" y="461" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="12" fill="#e8e6df">needle build --platform <tspan fill="#e8593d">macos-arm64</tspan> --layers 8 --out ./device</text>
+<text class="s c1" x="58" y="461" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="12" fill="#e8e6df">needle build --platform <tspan fill="#e8593d">linux-x86_64</tspan> --layers 8 --out ./device</text>
+<text class="s c2" x="58" y="461" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="12" fill="#e8e6df">needle build --platform <tspan fill="#e8593d">linux-arm64</tspan> --layers 8 --out ./device</text>
+<text class="s c3" x="58" y="461" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="12" fill="#e8e6df">needle build --platform <tspan fill="#e8593d">linux-armv7</tspan> --layers 8 --out ./device</text>
+<text class="s c4" x="58" y="461" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="12" fill="#e8e6df">needle build --platform <tspan fill="#e8593d">linux-riscv64</tspan> --layers 8 --out ./device</text>
+<text class="s c5" x="58" y="461" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="12" fill="#e8e6df">needle build --platform <tspan fill="#e8593d">linux-mipsel</tspan> --layers 8 --out ./device</text>
+<text class="s c6" x="58" y="461" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="12" fill="#e8e6df">needle build --platform <tspan fill="#e8593d">windows-x86_64</tspan> --layers 8 --out ./device</text>
+<text class="s c7" x="58" y="461" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="12" fill="#e8e6df">needle build --platform <tspan fill="#e8593d">windows-arm64</tspan> --layers 8 --out ./device</text>
+<text class="s c8" x="58" y="461" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="12" fill="#e8e6df">needle build --platform <tspan fill="#e8593d">android-arm64</tspan> --layers 8 --out ./device</text>
+<text class="s c9" x="58" y="461" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="12" fill="#e8e6df">needle build --platform <tspan fill="#e8593d">ios-arm64</tspan> --layers 8 --out ./device</text>
+<text class="s c10" x="58" y="461" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="12" fill="#e8e6df">needle build --platform <tspan fill="#e8593d">watchos-arm64</tspan> --layers 8 --out ./device</text>
+<text class="s c11" x="58" y="461" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="12" fill="#e8e6df">needle build --platform <tspan fill="#e8593d">wasm</tspan> --layers 8 --out ./device</text>
+<text class="s c12" x="58" y="461" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="12" fill="#e8e6df">needle build --platform <tspan fill="#e8593d">wasm-component</tspan> --layers 8 --out ./device</text>
+<text x="40" y="483" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10.5" fill="#4a4844">downloads the folder and writes needle3.cact beside the engine, at any depth from 2 to 20 layers</text>
+</svg>
\ No newline at end of file
diff --git a/assets/finetune.svg b/assets/finetune.svg
new file mode 100644
index 0000000..9dc03a5
--- /dev/null
+++ b/assets/finetune.svg
@@ -0,0 +1,85 @@
+<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 900 540" width="900" height="540" role="img" aria-label="Each Needle 3 subnetwork before and after fine-tuning on DroidCall and on Mobile Actions, against DeepSeek V4 Flash">
+<style>svg{--t0:0.0s}.s{animation-duration:13.0s;animation-iteration-count:infinite;animation-fill-mode:both;animation-delay:calc(0s - var(--t0))}@keyframes sc0{0%{opacity:0}0.00%{opacity:0}2.69%{opacity:1}47.31%{opacity:1}50.00%{opacity:0}100%{opacity:0}}.sc0{animation-name:sc0}@keyframes bar0{0%{transform:scaleY(0)}2.31%{transform:scaleY(0)}8.46%{transform:scaleY(1)}47.31%{transform:scaleY(1)}100%{transform:scaleY(0)}}.bar0{animation-name:bar0;transform-box:fill-box;transform-origin:center bottom;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes val0{0%{opacity:0}7.69%{opacity:0}10.77%{opacity:1}47.31%{opacity:1}100%{opacity:0}}.val0{animation-name:val0}@keyframes ref0{0%{transform:scaleX(0)}9.23%{transform:scaleX(0)}16.92%{transform:scaleX(1)}47.31%{transform:scaleX(1)}100%{transform:scaleX(0)}}.ref0{animation-name:ref0;transform-box:fill-box;transform-origin:left center;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes sc1{0%{opacity:0}50.00%{opacity:0}52.69%{opacity:1}97.31%{opacity:1}100.00%{opacity:0}100%{opacity:0}}.sc1{animation-name:sc1}@keyframes bar1{0%{transform:scaleY(0)}52.31%{transform:scaleY(0)}58.46%{transform:scaleY(1)}97.31%{transform:scaleY(1)}100%{transform:scaleY(0)}}.bar1{animation-name:bar1;transform-box:fill-box;transform-origin:center bottom;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes val1{0%{opacity:0}57.69%{opacity:0}60.77%{opacity:1}97.31%{opacity:1}100%{opacity:0}}.val1{animation-name:val1}@keyframes ref1{0%{transform:scaleX(0)}59.23%{transform:scaleX(0)}66.92%{transform:scaleX(1)}97.31%{transform:scaleX(1)}100%{transform:scaleX(0)}}.ref1{animation-name:ref1;transform-box:fill-box;transform-origin:left center;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@media (prefers-reduced-motion:reduce){.s{animation:none!important;opacity:1;transform:none}.sc1{display:none}}</style>
+<rect x="0" y="0" width="900" height="540" rx="18" fill="#0e0e10"/>
+<line x1="90" y1="332.5" x2="850" y2="332.5" stroke="#232326" stroke-width="1"/>
+<line x1="90" y1="245.0" x2="850" y2="245.0" stroke="#232326" stroke-width="1"/>
+<line x1="90" y1="157.5" x2="850" y2="157.5" stroke="#232326" stroke-width="1"/>
+<line x1="90" y1="70.0" x2="850" y2="70.0" stroke="#232326" stroke-width="1"/>
+<line x1="90" y1="420" x2="850" y2="420" stroke="#8a877f" stroke-width="1"/>
+<text x="78" y="424.0" text-anchor="end" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" fill="#85827a">0</text>
+<text x="78" y="336.5" text-anchor="end" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" fill="#85827a">25</text>
+<text x="78" y="249.0" text-anchor="end" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" fill="#85827a">50</text>
+<text x="78" y="161.5" text-anchor="end" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" fill="#85827a">75</text>
+<text x="78" y="74.0" text-anchor="end" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" fill="#85827a">100</text>
+<text x="30" y="245" text-anchor="middle" transform="rotate(-90 30 245)" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" fill="#85827a">accuracy (%)</text>
+<text x="166.0" y="444" text-anchor="middle" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" fill="#85827a">2L · 25M</text>
+<text x="318.0" y="444" text-anchor="middle" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" fill="#85827a">4L · 29M</text>
+<text x="470.0" y="444" text-anchor="middle" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" fill="#85827a">8L · 52M</text>
+<text x="622.0" y="444" text-anchor="middle" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" fill="#85827a">16L · 98M</text>
+<text x="774.0" y="444" text-anchor="middle" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" fill="#85827a">20L · 121M</text>
+<text x="450" y="520" text-anchor="middle" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10.5" fill="#6b6862">base and tuned scored with forced calls · DeepSeek V4 Flash through its cloud API</text>
+<g class="s sc0">
+<text x="90" y="40" font-family="Inter, -apple-system, Segoe UI, Helvetica, Arial, sans-serif" font-size="15" font-weight="600" fill="#e8e6df">Every subnetwork, fine-tuned on DroidCall</text>
+<text x="850" y="40" text-anchor="end" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" fill="#85827a">1 of 2</text>
+<line class="s ref0" x1="90" y1="208.2" x2="850" y2="208.2" stroke="#c9c6bd" stroke-width="1.2" stroke-dasharray="1 5" stroke-linecap="round"/>
+<text class="s val0" x="854" y="212.2" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10.5" fill="#c9c6bd">60.5</text>
+<rect class="s bar0" style="animation-delay:calc(0.00s - var(--t0))" x="122.0" y="348.2" width="40" height="71.8" fill="#6e6c66"/>
+<text class="s val0" style="animation-delay:calc(0.00s - var(--t0))" x="142.0" y="342.2" text-anchor="middle" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" font-weight="400" fill="#85827a">20.5</text>
+<rect class="s bar0" style="animation-delay:calc(0.06s - var(--t0))" x="170.0" y="222.3" width="40" height="197.7" fill="#e8593d"/>
+<text class="s val0" style="animation-delay:calc(0.06s - var(--t0))" x="190.0" y="216.3" text-anchor="middle" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" font-weight="600" fill="#e8e6df">56.5</text>
+<rect class="s bar0" style="animation-delay:calc(0.10s - var(--t0))" x="274.0" y="299.2" width="40" height="120.8" fill="#6e6c66"/>
+<text class="s val0" style="animation-delay:calc(0.10s - var(--t0))" x="294.0" y="293.2" text-anchor="middle" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" font-weight="400" fill="#85827a">34.5</text>
+<rect class="s bar0" style="animation-delay:calc(0.16s - var(--t0))" x="322.0" y="201.2" width="40" height="218.8" fill="#e8593d"/>
+<text class="s val0" style="animation-delay:calc(0.16s - var(--t0))" x="342.0" y="195.2" text-anchor="middle" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" font-weight="600" fill="#e8e6df">62.5</text>
+<rect class="s bar0" style="animation-delay:calc(0.20s - var(--t0))" x="426.0" y="271.2" width="40" height="148.8" fill="#6e6c66"/>
+<text class="s val0" style="animation-delay:calc(0.20s - var(--t0))" x="446.0" y="265.2" text-anchor="middle" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" font-weight="400" fill="#85827a">42.5</text>
+<rect class="s bar0" style="animation-delay:calc(0.26s - var(--t0))" x="474.0" y="182.0" width="40" height="238.0" fill="#e8593d"/>
+<text class="s val0" style="animation-delay:calc(0.26s - var(--t0))" x="494.0" y="176.0" text-anchor="middle" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" font-weight="600" fill="#e8e6df">68.0</text>
+<rect class="s bar0" style="animation-delay:calc(0.30s - var(--t0))" x="578.0" y="246.8" width="40" height="173.2" fill="#6e6c66"/>
+<text class="s val0" style="animation-delay:calc(0.30s - var(--t0))" x="598.0" y="240.8" text-anchor="middle" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" font-weight="400" fill="#85827a">49.5</text>
+<rect class="s bar0" style="animation-delay:calc(0.36s - var(--t0))" x="626.0" y="178.5" width="40" height="241.5" fill="#e8593d"/>
+<text class="s val0" style="animation-delay:calc(0.36s - var(--t0))" x="646.0" y="172.5" text-anchor="middle" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" font-weight="600" fill="#e8e6df">69.0</text>
+<rect class="s bar0" style="animation-delay:calc(0.40s - var(--t0))" x="730.0" y="238.0" width="40" height="182.0" fill="#6e6c66"/>
+<text class="s val0" style="animation-delay:calc(0.40s - var(--t0))" x="750.0" y="232.0" text-anchor="middle" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" font-weight="400" fill="#85827a">52.0</text>
+<rect class="s bar0" style="animation-delay:calc(0.46s - var(--t0))" x="778.0" y="175.0" width="40" height="245.0" fill="#e8593d"/>
+<text class="s val0" style="animation-delay:calc(0.46s - var(--t0))" x="798.0" y="169.0" text-anchor="middle" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" font-weight="600" fill="#e8e6df">70.0</text>
+<rect x="175.0" y="478" width="10" height="10" fill="#6e6c66"/>
+<text x="191.0" y="487" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" fill="#9a978f">Needle 3 base</text>
+<rect x="325.0" y="478" width="10" height="10" fill="#e8593d"/>
+<text x="341.0" y="487" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" fill="#9a978f">fine-tuned on DroidCall</text>
+<line x1="565.0" y1="483" x2="575.0" y2="483" stroke="#c9c6bd" stroke-width="1.2" stroke-dasharray="1 3" stroke-linecap="round"/>
+<text x="581.0" y="487" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" fill="#9a978f">DeepSeek V4 Flash</text>
+</g>
+<g class="s sc1">
+<text x="90" y="40" font-family="Inter, -apple-system, Segoe UI, Helvetica, Arial, sans-serif" font-size="15" font-weight="600" fill="#e8e6df">Every subnetwork, fine-tuned on Mobile Actions</text>
+<text x="850" y="40" text-anchor="end" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" fill="#85827a">2 of 2</text>
+<line class="s ref1" x1="90" y1="110.6" x2="850" y2="110.6" stroke="#c9c6bd" stroke-width="1.2" stroke-dasharray="1 5" stroke-linecap="round"/>
+<text class="s val1" x="854" y="114.6" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="10.5" fill="#c9c6bd">88.4</text>
+<rect class="s bar1" style="animation-delay:calc(0.00s - var(--t0))" x="122.0" y="416.9" width="40" height="3.1" fill="#6e6c66"/>
+<text class="s val1" style="animation-delay:calc(0.00s - var(--t0))" x="142.0" y="410.9" text-anchor="middle" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" font-weight="400" fill="#85827a">0.9</text>
+<rect class="s bar1" style="animation-delay:calc(0.06s - var(--t0))" x="170.0" y="186.5" width="40" height="233.5" fill="#e8593d"/>
+<text class="s val1" style="animation-delay:calc(0.06s - var(--t0))" x="190.0" y="180.5" text-anchor="middle" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" font-weight="600" fill="#e8e6df">66.7</text>
+<rect class="s bar1" style="animation-delay:calc(0.10s - var(--t0))" x="274.0" y="353.5" width="40" height="66.5" fill="#6e6c66"/>
+<text class="s val1" style="animation-delay:calc(0.10s - var(--t0))" x="294.0" y="347.5" text-anchor="middle" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" font-weight="400" fill="#85827a">19.0</text>
+<rect class="s bar1" style="animation-delay:calc(0.16s - var(--t0))" x="322.0" y="143.2" width="40" height="276.8" fill="#e8593d"/>
+<text class="s val1" style="animation-delay:calc(0.16s - var(--t0))" x="342.0" y="137.2" text-anchor="middle" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" font-weight="600" fill="#e8e6df">79.1</text>
+<rect class="s bar1" style="animation-delay:calc(0.20s - var(--t0))" x="426.0" y="261.8" width="40" height="158.2" fill="#6e6c66"/>
+<text class="s val1" style="animation-delay:calc(0.20s - var(--t0))" x="446.0" y="255.8" text-anchor="middle" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" font-weight="400" fill="#85827a">45.2</text>
+<rect class="s bar1" style="animation-delay:calc(0.26s - var(--t0))" x="474.0" y="127.0" width="40" height="293.0" fill="#e8593d"/>
+<text class="s val1" style="animation-delay:calc(0.26s - var(--t0))" x="494.0" y="121.0" text-anchor="middle" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" font-weight="600" fill="#e8e6df">83.7</text>
+<rect class="s bar1" style="animation-delay:calc(0.30s - var(--t0))" x="578.0" y="139.3" width="40" height="280.7" fill="#6e6c66"/>
+<text class="s val1" style="animation-delay:calc(0.30s - var(--t0))" x="598.0" y="133.3" text-anchor="middle" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" font-weight="400" fill="#85827a">80.2</text>
+<rect class="s bar1" style="animation-delay:calc(0.36s - var(--t0))" x="626.0" y="123.9" width="40" height="296.1" fill="#e8593d"/>
+<text class="s val1" style="animation-delay:calc(0.36s - var(--t0))" x="646.0" y="117.9" text-anchor="middle" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" font-weight="600" fill="#e8e6df">84.6</text>
+<rect class="s bar1" style="animation-delay:calc(0.40s - var(--t0))" x="730.0" y="116.6" width="40" height="303.4" fill="#6e6c66"/>
+<text class="s val1" style="animation-delay:calc(0.40s - var(--t0))" x="750.0" y="110.6" text-anchor="middle" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" font-weight="400" fill="#85827a">86.7</text>
+<rect class="s bar1" style="animation-delay:calc(0.46s - var(--t0))" x="778.0" y="124.2" width="40" height="295.8" fill="#e8593d"/>
+<text class="s val1" style="animation-delay:calc(0.46s - var(--t0))" x="798.0" y="118.2" text-anchor="middle" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" font-weight="600" fill="#e8e6df">84.5</text>
+<rect x="175.0" y="478" width="10" height="10" fill="#6e6c66"/>
+<text x="191.0" y="487" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" fill="#9a978f">Needle 3 base</text>
+<rect x="325.0" y="478" width="10" height="10" fill="#e8593d"/>
+<text x="341.0" y="487" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" fill="#9a978f">fine-tuned on Mobile Actions</text>
+<line x1="565.0" y1="483" x2="575.0" y2="483" stroke="#c9c6bd" stroke-width="1.2" stroke-dasharray="1 3" stroke-linecap="round"/>
+<text x="581.0" y="487" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" fill="#9a978f">DeepSeek V4 Flash</text>
+</g>
+</svg>
\ No newline at end of file
diff --git a/assets/frontier.png b/assets/frontier.png
deleted file mode 100644
index 8e79d6c..0000000
Binary files a/assets/frontier.png and /dev/null differ
diff --git a/assets/frontier.svg b/assets/frontier.svg
deleted file mode 100644
index f00812c..0000000
--- a/assets/frontier.svg
+++ /dev/null
@@ -1,42 +0,0 @@
-<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 900 520" role="img" aria-label="Scatter plot of Mobile Actions accuracy versus model size">
- <rect x="0" y="0" width="900" height="520" fill="#111113"/>
- <line x1="90" y1="60" x2="850" y2="60" stroke="#232326" stroke-width="1"/>
- <line x1="90" y1="150" x2="850" y2="150" stroke="#232326" stroke-width="1"/>
- <line x1="90" y1="240" x2="850" y2="240" stroke="#232326" stroke-width="1"/>
- <line x1="90" y1="330" x2="850" y2="330" stroke="#232326" stroke-width="1"/>
- <line x1="90" y1="420" x2="850" y2="420" stroke="#8a877f" stroke-width="1"/>
- <line x1="90" y1="60" x2="90" y2="420" stroke="#8a877f" stroke-width="1"/>
- <g font-family="ui-monospace, Menlo, monospace" font-size="11" fill="#85827a">
- <text x="78" y="424" text-anchor="end">40</text>
- <text x="78" y="334" text-anchor="end">55</text>
- <text x="78" y="244" text-anchor="end">70</text>
- <text x="78" y="154" text-anchor="end">85</text>
- <text x="78" y="64" text-anchor="end">100</text>
- <text x="90" y="444" text-anchor="middle">0</text>
- <text x="209" y="444" text-anchor="middle">0.5B</text>
- <text x="328" y="444" text-anchor="middle">1B</text>
- <text x="446" y="444" text-anchor="middle">1.5B</text>
- <text x="565" y="444" text-anchor="middle">2B</text>
- <text x="684" y="444" text-anchor="middle">2.5B</text>
- <text x="802" y="444" text-anchor="middle">3B</text>
- <text x="470" y="480" text-anchor="middle">Total parameters</text>
- <text x="30" y="240" text-anchor="middle" transform="rotate(-90 30 240)">Mobile-Actions accuracy (%)</text>
- </g>
- <path d="M 144.6 245.4 C 149.3 245.4, 149.3 276.0, 154.1 276.0 C 478.3 276.0, 478.3 314.4, 802.5 314.4" fill="none" stroke="#8a877f" stroke-width="1.4"/>
- <g font-family="Inter, -apple-system, sans-serif" font-size="12.5" fill="#e8e6df">
- <circle cx="144.6" cy="245.4" r="5" fill="#e8e6df"/>
- <text x="156.6" y="237.4">LFM2.5 230M</text>
- <text x="156.6" y="250.4" font-size="10.5" fill="#85827a">f16 · vLLM</text>
- <circle cx="154.1" cy="276.0" r="5" fill="#e8e6df"/>
- <text x="166.1" y="292.0">FunctionGemma 270M</text>
- <text x="166.1" y="305.0" font-size="10.5" fill="#85827a">f16 · vLLM</text>
- <circle cx="802.5" cy="314.4" r="5" fill="#e8e6df"/>
- <text x="790.5" y="334.4" text-anchor="end">Apple FM</text>
- <text x="790.5" y="347.4" text-anchor="end" font-size="10.5" fill="#85827a">on-device</text>
- </g>
- <g>
- <circle cx="100.7" cy="277.8" r="6" fill="#e8593d"/>
- <text x="100.7" y="251.8" text-anchor="middle" font-family="Inter, sans-serif" font-size="13" font-weight="600" fill="#e8593d">Needle 2</text>
- <text x="100.7" y="264.8" text-anchor="middle" font-size="10.5" fill="#85827a">CQ2-bit</text>
- </g>
- </svg>
diff --git a/assets/model.svg b/assets/model.svg
new file mode 100644
index 0000000..4a203bf
--- /dev/null
+++ b/assets/model.svg
@@ -0,0 +1,37 @@
+<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 900 346" width="900" height="346" role="img" aria-label="Needle 3 at a glance: inputs, outputs, model, training, speed and customisation">
+<style>svg{--t0:0.0s}.s{animation-duration:9.0s;animation-iteration-count:infinite;animation-fill-mode:both;animation-delay:calc(0s - var(--t0))}@keyframes r0{0%{opacity:0;transform:translateX(-10px)}3.33%{opacity:0;transform:translateX(-10px)}8.33%{opacity:1;transform:translateX(0)}93.33%{opacity:1;transform:translateX(0)}98.89%{opacity:0}100%{opacity:0;transform:translateX(-10px)}}.r0{animation-name:r0;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes l0{0%{transform:scaleX(0)}3.33%{transform:scaleX(0)}10.00%{transform:scaleX(1)}93.33%{transform:scaleX(1)}98.89%{transform:scaleX(1);opacity:0}100%{transform:scaleX(0);opacity:1}}.l0{animation-name:l0;transform-box:fill-box;transform-origin:left center;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes r1{0%{opacity:0;transform:translateX(-10px)}9.44%{opacity:0;transform:translateX(-10px)}14.44%{opacity:1;transform:translateX(0)}93.33%{opacity:1;transform:translateX(0)}98.89%{opacity:0}100%{opacity:0;transform:translateX(-10px)}}.r1{animation-name:r1;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes l1{0%{transform:scaleX(0)}9.44%{transform:scaleX(0)}16.11%{transform:scaleX(1)}93.33%{transform:scaleX(1)}98.89%{transform:scaleX(1);opacity:0}100%{transform:scaleX(0);opacity:1}}.l1{animation-name:l1;transform-box:fill-box;transform-origin:left center;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes r2{0%{opacity:0;transform:translateX(-10px)}15.56%{opacity:0;transform:translateX(-10px)}20.56%{opacity:1;transform:translateX(0)}93.33%{opacity:1;transform:translateX(0)}98.89%{opacity:0}100%{opacity:0;transform:translateX(-10px)}}.r2{animation-name:r2;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes l2{0%{transform:scaleX(0)}15.56%{transform:scaleX(0)}22.22%{transform:scaleX(1)}93.33%{transform:scaleX(1)}98.89%{transform:scaleX(1);opacity:0}100%{transform:scaleX(0);opacity:1}}.l2{animation-name:l2;transform-box:fill-box;transform-origin:left center;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes r3{0%{opacity:0;transform:translateX(-10px)}21.67%{opacity:0;transform:translateX(-10px)}26.67%{opacity:1;transform:translateX(0)}93.33%{opacity:1;transform:translateX(0)}98.89%{opacity:0}100%{opacity:0;transform:translateX(-10px)}}.r3{animation-name:r3;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes l3{0%{transform:scaleX(0)}21.67%{transform:scaleX(0)}28.33%{transform:scaleX(1)}93.33%{transform:scaleX(1)}98.89%{transform:scaleX(1);opacity:0}100%{transform:scaleX(0);opacity:1}}.l3{animation-name:l3;transform-box:fill-box;transform-origin:left center;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes r4{0%{opacity:0;transform:translateX(-10px)}27.78%{opacity:0;transform:translateX(-10px)}32.78%{opacity:1;transform:translateX(0)}93.33%{opacity:1;transform:translateX(0)}98.89%{opacity:0}100%{opacity:0;transform:translateX(-10px)}}.r4{animation-name:r4;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes l4{0%{transform:scaleX(0)}27.78%{transform:scaleX(0)}34.44%{transform:scaleX(1)}93.33%{transform:scaleX(1)}98.89%{transform:scaleX(1);opacity:0}100%{transform:scaleX(0);opacity:1}}.l4{animation-name:l4;transform-box:fill-box;transform-origin:left center;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes r5{0%{opacity:0;transform:translateX(-10px)}33.89%{opacity:0;transform:translateX(-10px)}38.89%{opacity:1;transform:translateX(0)}93.33%{opacity:1;transform:translateX(0)}98.89%{opacity:0}100%{opacity:0;transform:translateX(-10px)}}.r5{animation-name:r5;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@keyframes l5{0%{transform:scaleX(0)}33.89%{transform:scaleX(0)}40.56%{transform:scaleX(1)}93.33%{transform:scaleX(1)}98.89%{transform:scaleX(1);opacity:0}100%{transform:scaleX(0);opacity:1}}.l5{animation-name:l5;transform-box:fill-box;transform-origin:left center;animation-timing-function:cubic-bezier(.2,.8,.2,1)}@media (prefers-reduced-motion:reduce){.s{animation:none!important;opacity:1;transform:none}}</style>
+<rect width="900" height="346" rx="18" fill="#0e0e10"/>
+<text x="40" y="38" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="15" font-weight="600" fill="#e8e6df">Needle 3 at a glance</text>
+<line class="s l0" x1="40" y1="66" x2="860" y2="66" stroke="#232326" stroke-width="1"/>
+<g class="s r0">
+<text x="40" y="90" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" letter-spacing="1.5" fill="#e8593d">INPUTS</text>
+<text x="210" y="90" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="15" fill="#e8e6df">Text prompts, plus tool definitions or an extraction schema</text>
+</g>
+<line class="s l1" x1="40" y1="104" x2="860" y2="104" stroke="#232326" stroke-width="1"/>
+<g class="s r1">
+<text x="40" y="128" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" letter-spacing="1.5" fill="#e8593d">OUTPUTS</text>
+<text x="210" y="128" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="15" fill="#e8e6df">Tool calls or structured extraction, always well-formed</text>
+</g>
+<line class="s l2" x1="40" y1="142" x2="860" y2="142" stroke="#232326" stroke-width="1"/>
+<g class="s r2">
+<text x="40" y="166" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" letter-spacing="1.5" fill="#e8593d">MODEL</text>
+<text x="210" y="166" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="15" fill="#e8e6df">29-121M Laddered Simple Attention Networks, CQ2 quantisation</text>
+</g>
+<line class="s l3" x1="40" y1="180" x2="860" y2="180" stroke="#232326" stroke-width="1"/>
+<g class="s r3">
+<text x="40" y="204" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" letter-spacing="1.5" fill="#e8593d">TRAINING</text>
+<text x="210" y="204" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="15" fill="#e8e6df">360B tokens of proprietary structured dataset</text>
+</g>
+<line class="s l4" x1="40" y1="218" x2="860" y2="218" stroke="#232326" stroke-width="1"/>
+<g class="s r4">
+<text x="40" y="242" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" letter-spacing="1.5" fill="#e8593d">SPEED</text>
+<text x="210" y="242" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="15" fill="#e8e6df">Up to 4k tokens/s decode and 10k tokens/s prefill on a Raspberry Pi 5</text>
+</g>
+<line class="s l5" x1="40" y1="256" x2="860" y2="256" stroke="#232326" stroke-width="1"/>
+<g class="s r5">
+<text x="40" y="280" font-family="ui-monospace, SFMono-Regular, Menlo, Consolas, monospace" font-size="11" letter-spacing="1.5" fill="#e8593d">CUSTOMISATION</text>
+<text x="210" y="280" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="15" fill="#e8e6df">Fine-tuning lifts every subnetwork 18 to 36 points on DroidCall;</text>
+<text x="210" y="302" font-family="Inter, -apple-system, 'Segoe UI', Helvetica, Arial, sans-serif" font-size="15" fill="#e8e6df">from 4 layers (29M) up the tuned subnetwork passes DeepSeek V4 Flash</text>
+</g>
+<line x1="40" y1="316" x2="860" y2="316" stroke="#232326" stroke-width="1"/>
+</svg>
\ No newline at end of file
diff --git a/doc/apis.md b/doc/apis.md
deleted file mode 100644
index dfe2a07..0000000
--- a/doc/apis.md
+++ /dev/null
@@ -1,186 +0,0 @@
-# API
-
-| Call | What it does |
-| --- | --- |
-| `needle.Needle(tools=None, system=None, weights=None, tool_index_path=None, buffer_size=65536)` | Create an agent bound to one toolset. `tools` takes decorated functions, Pydantic models, raw JSON schema dicts, or a JSON string. `system` carries environment facts. `weights` loads a tuned `.cact`. `tool_index_path` persists tool embeddings for large catalogues. |
-| `agent.complete(text="", max_new_tokens=512)` | One turn. You execute the call and feed the result back yourself via the next `complete(...)`. |
-| `agent.embed(text="")` | Embed text or serialized tool schemas with the Needle 3 retrieval head. |
-| `agent.run(query="", max_steps=8, max_new_tokens=512, strict=True)` | Full agentic loop: the model picks calls, Needle executes your Python functions and feeds results back, and the final response carries the executed tool results as `results`. |
-| `agent.reset()` | Rewind the conversation, keep the tools loaded. |
-| `needle.tool` | Decorator that turns a function into a tool schema (attached as `fn._needle_tool`). |
-| `needle.Field(...)` | Per argument constraints, attached inline with `typing.Annotated` or passed as a default. |
-| `needle.extract(text, schema, system=None, max_new_tokens=512, weights=None, strict=True)` | One shot extraction. Returns a Pydantic instance if `schema` is a model, else a dict, or `None` if nothing matched. Strict mode rejects temporal values that contradict literal source years and engine-reported ungrounded values. |
-
-When one tuned archive is active, `extract(..., weights=None)` inherits it for backwards compatibility. If both a Needle 2 and Needle 3 tune are active, pass `weights=` explicitly; the client refuses to guess between generations.
-
-## Declaring tools
-
-**Describe each argument and offer choices.** Needle reads a Google-style `Args:` block for per-parameter descriptions; a default makes an argument optional; a `Literal` becomes a fixed set the model must choose from (it cannot emit anything else).
-
-```python
-from typing import Literal
-
-def set_thermostat(temperature: int, mode: Literal["heat", "cool", "auto"] = "auto"):
- """Set the thermostat.
-
- Args:
- temperature: target temperature in Celsius
- mode: heating strategy to use
- """
- return {"temperature": temperature, "mode": mode}
-
-agent = needle.Needle(tools=[set_thermostat])
-agent.run("make it 21 and cool the room")
-```
-
-**Constrain the values** with `needle.Field`, attached inline via `Annotated`. Ranges, patterns, lengths, and item counts are compiled into the decode grammar, so the model can only ever emit values that satisfy them.
-
-```python
-from typing import Annotated
-
-def send_money(
- amount: Annotated[float, needle.Field(gt=0, le=10000, description="USD, up to 10,000")],
- to: Annotated[str, needle.Field(pattern=r"^@[a-z0-9_]+$", description="recipient handle")],
- memo: Annotated[str, needle.Field(max_length=80)] = "",
-):
- "Send money to a handle."
- return {"sent": amount, "to": to}
-```
-
-`Field` supports `description`, `enum`, `const`, `ge`/`le`/`gt`/`lt`, `multiple_of`, `min_length`/`max_length`, `pattern`, `format`, `min_items`/`max_items`, and `unique_items`.
-
-**By hand.** The decorator just builds a JSON schema; you can pass that schema directly, which is exactly what Needle consumes. This is how you set descriptions and constraints without the decorator:
-
-```python
-tools = [{
- "name": "set_lights",
- "description": "Turn a room's lights on or off and set brightness",
- "parameters": {
- "type": "object",
- "properties": {
- "room": {"type": "string", "description": "which room to control"},
- "on": {"type": "boolean"},
- "brightness": {"type": "integer", "minimum": 0, "maximum": 100},
- },
- "required": ["room", "on"],
- },
-}]
-agent = needle.Needle(tools=tools)
-```
-
-## Driving the loop
-
-Prefer to drive the loop yourself instead of `run()`? `complete()` returns the raw call and you execute it:
-
-```python
-import json
-response = agent.complete("dim the living room to 30")
-if response["type"] == "call":
- result = set_lights(**response["function_calls"][0]["arguments"])
- response = agent.complete(json.dumps(result)) # feed the result back
-```
-
-Every turn returns one JSON object:
-
-```json
-{
- "type": "call",
- "success": true,
- "error": null,
- "error_code": null,
- "function_calls": [ { "name": "set_lights", "arguments": { "room": "living room", "on": true, "brightness": 30 } } ],
- "reasoning": "'living room' -> room; 'dim' -> on true, brightness 30",
- "confidence": 0.94,
- "prefill_tps": 4300.0,
- "decode_tps": 850.0,
- "peak_ram_mb": 28.5
-}
-```
-
-## Behaviour
-
-Needle solves every problem as a function call. The context declares what may be called; the model answers with calls. Performing an action and extracting structured data are the same operation, the only difference is what you declare.
-
-- A request no declared tool can serve is refused with the empty call `[]`. That is the whole contract for off-topic input; there is no free-text fallback.
-- Arguments contain only values evidenced by the input. An optional field with no evidence is omitted, not guessed; omission is the field-level `[]`.
-- A call below the confidence threshold, or contradicted by the request, is withheld by the engine: `function_calls` is empty and `suppressed_calls` carries it. The agent prefixes the local `date:` fact to `system` automatically (`auto_date=False` opts out) so relative dates resolve.
-- A date argument whose year matches none of the years written in the conversation so far or in the `system` facts is reported in `validation.ungrounded` as `tool.field`, alongside anything the engine itself flags as ungrounded. `run()` does not execute such a call; its result is `{"error": "ungrounded field"}` and the model continues from it. Pass `strict=False` to execute anyway.
-- `extract()` declares the record as the only tool and takes the record from `function_calls` or, when the engine withheld it, from `suppressed_calls`; `strict` handles grounding.
-- Arguments arrive after the engine's deterministic repair step, which grounds them in the request text (first/last name split, phone completion, verbatim quoted titles, place queries with their street address, weekday and explicit dates resolved against the `date:` fact, empty optionals and duplicate calls dropped). Enumerated parameters are never altered.
-- `reasoning` is the model's short derivation of each argument from its source span (`'ten minutes' -> minutes 10`). It is generated unconstrained; only the call itself is grammar-constrained, so the JSON cannot be malformed while the derivation stays legible.
-- After you execute a call, pass the result back as the next `complete()`. The model continues from it, and later arguments may depend on earlier results: `search_for_contact` first, then `send_instant_message` with the returned `contact_id`. A final `"type": "respond"` with empty `function_calls` signals the loop is done; the answer is the tool results themselves, which `run()` collects on the final response as `results`. No free text is generated.
-- An agent shares one toolset. Later turns are bare queries against the same tools; `reset()` rewinds the conversation and keeps the tools loaded.
-- Each tuned agent runs in its own worker process. The worker loads its `.cact` once and owns an independent engine, KV cache, and conversation without extending the native C API.
-- Needle 2 and Needle 3 archives carry different format tags. The package reads the tag and loads the matching native engine; it never feeds one generation's weights to the other engine.
-
-## Extraction
-
-Extraction is not a separate mode - it is tool calling with one tool. Declare the record as the only schema and pass the content where the query goes; the returned call's `arguments` are the extracted fields. With one declared tool the grammar admits exactly one call of that name, so schema conformance is guaranteed rather than requested. Use the `extract()` helper for a typed result, or pass a plain schema and read the call:
-
-```python
-receipt = [{
- "name": "receipt",
- "description": "A purchase receipt shared as text",
- "parameters": {
- "type": "object",
- "properties": {
- "merchant": {"type": "string"},
- "total": {"type": "number"},
- "currency": {"type": "string"},
- "line_items": {"type": "array", "items": {"type": "object"}},
- },
- "required": ["merchant", "total"],
- },
-}]
-agent = needle.Needle(tools=receipt)
-print(agent.complete("GreenMart receipt: oat milk 3.50, total 7.75 paid by visa")["function_calls"])
-# -> [{'name': 'receipt', 'arguments': {'merchant': 'GreenMart', 'total': 7.75}}]
-```
-
-Because it is the same operation, everything else applies unchanged: `confidence` gates the extraction, unsupported input returns the empty call `[]`, and fine-tuning uses the same data format (the record as the tool, the passage as the query).
-
-## System facts
-
-An optional system turn carries environment state as facts, never instructions:
-
-```python
-agent = needle.Needle(tools=tools, system="date: 2026-07-21 Tue 14:30; locale: en-US; device: phone; battery: 62%")
-```
-
-Recognized keys are `date`, `locale`, `device`, `battery`, `network`, `location`, `user`, and `assistant`. The model resolves relative language against them: "tomorrow at 7" becomes an absolute time only when a `date:` fact licenses it, otherwise the human phrase passes through verbatim. `assistant:` declares the identity the model binds to. Needle trains with and without the turn, so omitting it is safe; instructions placed there do not steer the model.
-
-## Tool retrieval
-
-Five or fewer declared tools render directly. Above that, retrieval engages: at init every tool schema is embedded once by a built-in contrastive head, each turn embeds the query, and only the five highest-scoring tools enter the context, with the grammar rebuilt over just that subset. An unselected tool is unreachable, not merely unlikely. `tool_index_path` persists the embeddings on disk, keyed by a fingerprint over the schemas and the model; a matching fingerprint loads instantly, a changed schema re-embeds only what changed.
-
-## Confidence
-
-The `confidence` field is the minimum of two signals: a calibrated post-hoc head that scores the full prompt plus the call the model just produced, and the decoding probability of the call tokens. A call is accepted only when both agree, so the failure mode is escalation, not wrong execution. The contract: pick a threshold for your product, act at or above it, re-ask or route to a bigger model below it. Off-topic requests return the empty call `[]`. A call below the engine's default threshold of 0.1 is withheld into `suppressed_calls`.
-
-Calibration holds for the base model only. Fine-tuning does not update the head, so an agent running tuned weights reports `confidence` as `None` and warns once at construction.
-
-## Offline devices
-
-Each generation's engine is fetched once and cached under `~/.cache/cactus-needle/v2/` or `v3/`; Needle 3 also caches its base archive `needle3.cact` there, since the engine carries no weights (`needle download needle3` fetches it explicitly). Inference itself never touches the network, so an air gapped device only needs the matching files in place. Three ways to get them there:
-
-1. `needle fetch --generation 2|3` downloads that engine for the current machine into the cache and prints the path. `--out <dir>` places it elsewhere. `--platform-tag manylinux2014_aarch64` fetches the build for a different device; tags follow the wheel names on the Hugging Face repo (`macosx_11_0_arm64`, `manylinux2014_x86_64`, `musllinux_1_2_aarch64`, `win_amd64`, `win_arm64`). For the standalone engine runner, `needle download <platform> --generation 2|3 [--out <dir>]` (platform folders include `macos-arm64`, `linux-x86_64`, `wasm`, and `wasm-component`; the CLI prints the full list on an invalid name) copies that platform's engine files into `<out>/<platform>/` and marks a native `needle` runner executable when present. The `wasm-component` target contains `needle.component.wasm` and the `cactus:needle/[email protected]` WIT contract.
-2. Copy the file to the same cache path on the device, or drop it inside the installed `needle/` package directory, which wins over the cache.
-3. Set `NEEDLE2_LIB_PATH=/path/to/libneedle.so` or `NEEDLE3_LIB_PATH=/path/to/libneedle.so` to override one generation. The legacy `NEEDLE_LIB_PATH` remains a Needle 2 alias; it is deliberately ignored for Needle 3 so a v3 archive cannot be sent to a v2 engine.
-
-The Python package itself installs offline the standard way: `pip download cactus-needle` on a connected machine, then `pip install --no-index --find-links <dir> cactus-needle` on the device. On a device that must never attempt the network, also set `HF_HUB_OFFLINE=1` so a missing engine fails fast with a clear error instead of trying to download.
-
-## WebAssembly component
-
-`needle download wasm-component --generation 2 --out .` pulls `needle.component.wasm` and its `needle.wit` world from the Hub. The same files are published to GHCR as a signed OCI artifact, one image per engine generation, tagged with the engine version:
-
-```sh
-oras pull ghcr.io/cactus-compute/needle2-component:2.0.4
-
-cosign verify ghcr.io/cactus-compute/needle2-component:2.0.4 \
- --certificate-identity-regexp '^https://github.com/cactus-compute/needle/' \
- --certificate-oidc-issuer https://token.actions.githubusercontent.com
-```
-
-The artifact type is `application/vnd.wasm.component.v1+wasm`, so `wkg oci pull` works as well. Signatures are Sigstore keyless from the `publish-component` workflow in this repository; the certificate identity above is the only trust anchor.
diff --git a/doc/environments.md b/doc/environments.md
deleted file mode 100644
index a7dbeea..0000000
--- a/doc/environments.md
+++ /dev/null
@@ -1,118 +0,0 @@
-# Environments
-
-Pre-configured tool surfaces in `needle.environments`. Each module is a hand-curated set of at most five tools whose enums, bounds, and descriptions map cleanly onto Needle's constrained decoding, plus a ready agent and a frozen acceptance suite.
-
-```python
-from needle.environments import smart_home
-
-smart_home.agent.complete("dim the study lights to 30 percent")
-smart_home.run_tests()
-```
-
-```sh
-python -m needle.environments.smart_home # run one frozen suite, exit 0 on pass
-```
-
-## Module contract
-
-Every environment module exposes the same surface:
-
-| Attribute | What it is |
-| --- | --- |
-| `TOOLS` | The decorated tool functions, each returning `{"ok": True, ...}` echoes of its arguments. |
-| `SYSTEM` | The system prompt the agent is built with. |
-| `agent` | A `needle.Needle(tools=TOOLS, system=SYSTEM)`, constructed lazily on first access so importing never fetches the engine. `run_tests` treats a call whose `validation` reports ungrounded values or a negation as a refusal. |
-| `TEST_CASES` | The frozen suite: dicts of `query`, expected `calls`, `category`, and optional `critical`. |
-| `run_tests(min_confidence=0.0, verbose=True)` | Runs the suite against the shipped engine; returns `True` at >=90% pass with zero critical failures. |
-
-`needle.environments.ENVIRONMENTS` maps name to module; `needle.environments.run_tests()` runs all six.
-
-## Available environments
-
-### smart_home
-
-Lights, thermostat, fans, blinds, robot vacuum, over four rooms (`kitchen`, `living_room`, `bedroom`, `study`).
-
-| Tool | Arguments |
-| --- | --- |
-| `control_lights` | `room`, `action` on/off/dim, `brightness_percent` 0-100, `color` from 5 names |
-| `set_thermostat` | `temperature` 10-30 C |
-| `control_fan` | `room` (no kitchen), `action` on/off, `speed` low/medium/high |
-| `control_blinds` | `room`, `action` open/close |
-| `start_robot_vacuum` | `action` start/stop/dock, `room` (no study) |
-
-### media_player
-
-Smart speaker playback and volume. `play_music` starts named content; the zero-argument playback tools control whatever is already playing.
-
-| Tool | Arguments |
-| --- | --- |
-| `play_music` | `query` verbatim, 1-80 chars |
-| `pause_media`, `resume_media`, `skip_track` | none |
-| `set_volume` | `level` 0-100 |
-
-### productivity
-
-Timers, reminders, calendar events, tasks, notes. Date and time phrases are copied verbatim for the host app to resolve; a reminder without a time is treated as incomplete, and undated to-dos belong to `add_task`.
-
-| Tool | Arguments |
-| --- | --- |
-| `set_timer` | `time_human` verbatim, 1-60 chars |
-| `create_reminder` | `message` 1-120 chars, `date_time_human` |
-| `create_calendar_event` | `title` 1-120 chars, `start_time_human`, `location` 1-80 chars |
-| `add_task` | `title` 1-120 chars, `priority` low/medium/high |
-| `create_note` | `text` 1-200 chars, `title` 1-60 chars |
-
-### wearable
-
-Watch notifications, workouts, find-my-phone. Replies only go to the four named senders (`Maya`, `Leo`, `Dr. Patel`, `Cactus Team`); workout types use gerund forms.
-
-| Tool | Arguments |
-| --- | --- |
-| `reply_to_notification` | `notification_match` sender enum, `text` verbatim, 1-240 chars |
-| `dismiss_notification` | `notification_match` sender enum |
-| `start_workout` | `workout_type` running/walking/cycling/swimming/strength/yoga |
-| `end_workout`, `find_my_phone` | none |
-
-### kitchen_appliance
-
-Oven, coffee maker, dishwasher, cooking timers. Hard numeric bounds make unsafe requests unrepresentable; the one read-only tool keeps checks separate from actions.
-
-| Tool | Arguments |
-| --- | --- |
-| `set_oven` | `temperature` 50-250 C |
-| `control_coffee_maker` | `action` brew/stop/warm |
-| `start_dishwasher` | `cycle` eco/heavy/quick |
-| `set_cooking_timer` | `label` 1-40 chars, `minutes` 1-360 |
-| `get_appliance_status` | `appliance` oven/coffee_maker/dishwasher, read-only |
-
-### data_capture
-
-Contacts, expenses, meals, water, weight: dictated facts into typed records. Every value is copied verbatim, a member of a closed set, or a bounded number; the phone pattern and email format are enforced during decoding.
-
-| Tool | Arguments |
-| --- | --- |
-| `create_contact` | `name` 1-60 chars, `phone` pattern `^\+?[0-9][0-9 -]{5,17}$`, `email` format email |
-| `log_expense` | `amount` 0-100000, `category` from 6, `merchant` 1-60 chars |
-| `log_meal` | `description` 1-120 chars, `meal_type` breakfast/lunch/dinner/snack |
-| `log_water_intake` | `amount_ml` 1-5000 |
-| `log_weight` | `weight_kg` 20-300 |
-
-## The frozen suites
-
-Each suite has 32 cases across six categories:
-
-| Category | Expectation |
-| --- | --- |
-| `positive` | The exact call, arguments included. |
-| `missing` | `[]`: a required value was not stated, do not guess it. |
-| `irrelevant` | `[]`: no declared tool covers the request. |
-| `negation` | `[]`: the request is negated ("don't", "do not", "never"). |
-| `invalid` | `[]`: a stated value is outside the declared bounds. |
-| `parallel` | Two calls from one query, order-insensitive. |
-
-Cases marked `critical: True` (all `missing`, `negation`, and `invalid` cases) fail the suite regardless of the overall rate. `run_tests()` scores raw model output; `run_tests(min_confidence=0.4)` applies the production contract, acting on a call only at or above the threshold and treating anything below as a refusal.
-
-## Adapting one
-
-Swap the `Literal` values (rooms, contacts, categories) for your own and keep the shapes: closed sets as enums, bounded numbers, verbatim copy for free text, five tools or fewer. One learned rule from smart_home: avoid enum values that hide inside likely query words (a room named office poisons an off action, so that home has a study).
diff --git a/doc/finetuning.md b/doc/finetuning.md
deleted file mode 100644
index 423a495..0000000
--- a/doc/finetuning.md
+++ /dev/null
@@ -1,108 +0,0 @@
-# Finetuning
-
-Needle finetunes with LoRA adapters on the frozen base model: rank 16 on the five attention projections of every layer, trained on your JSONL, then merged into the weights at export. The engine, the tokenizer, and the confidence head are untouched. The output is a `.cact` archive you pass as `weights=`.
-
-## Data format
-
-Install the training dependencies before running the commands in this guide:
-
-```sh
-pip install "cactus-needle[train]"
-```
-
-One JSON object per line. `query` and `tools` describe the turn, `answers` lists the exact calls the model should emit, `reasoning` is one short line deriving each argument from its source span in the query. `reasoning` is optional but include it: the model produces the derivation before the call, and examples that show where each value comes from teach grounding, not just tool selection.
-
-```json
-{"query": "Bantilan, N. (2018). Themis. Journal of Technology in Human Services, 36(1).", "tools": [{"name": "extract_citation_data", "parameters": {"type": "object", "properties": {"authors": {"type": "string"}, "title": {"type": "string"}, "publisher": {"type": "string"}}, "required": ["authors", "title"]}}], "answers": [{"name": "extract_citation_data", "arguments": {"authors": "Bantilan, N.", "title": "Themis", "publisher": "Journal of Technology in Human Services, 36(1)."}}], "reasoning": "authors precede the year; title follows the year; publisher is the journal segment"}
-```
-
-Rules that matter:
-
-1. Arguments contain only values present in the query. Omit optional fields with no evidence; never fill them with placeholders or empty strings.
-2. Include off topic examples with `"answers": []`. The built in generator produces about 1 in 8. Without them the tuned model calls a tool on everything.
-3. When the catalogue has similar tools, include ambiguous queries resolved to the correct one.
-4. An optional `"system"` field per example becomes a system turn, matching `Needle(system=...)` at inference.
-5. Each rendered example must fit within `--max-len` (default 1024) tokens; longer examples are silently truncated. Padding adjusts to the longest example automatically (rounded up to a power of two), so a short dataset trains fast regardless of the cap. This matters on CPU: a dataset of 245 token examples trains 6 times faster padded to 256 than to 1024.
-
-## Commands
-
-Train, export, load:
-
-```sh
-needle finetune data.jsonl --epochs 10 --out adapter.safetensors
-needle build checkpoints/needle3.safetensors --lora adapter.safetensors --out tuned.cact
-```
-
-Checkpoints and adapters are `.safetensors`; pickle files still load. The base is `checkpoints/needle3.safetensors`, the 20-layer Needle 3.
-
-```python
-agent = needle.Needle(tools=[...], weights="tuned.cact")
-```
-
-Fine-tuning is quantisation-aware: LoRA trains through the 4-bit CQ numerics
-`needle build` exports, at the full 20 layers of the base. Every build fetches the
-published `needle3.cact` and packages its tokenizer.
-
-## Depth
-
-Needle 3 is a depth ladder: every rung from 2 to 20 layers is a trained model whose
-blocks are a nested subset of the full stack. `needle build --layers n` merges the
-adapter into the full base and exports the n-layer rung, so one fine-tune serves
-every device budget.
-
-```sh
-needle build checkpoints/needle3.safetensors --lora adapter.safetensors --layers 8 --out tuned_8l.cact
-```
-
-To share a tuned model, set `NEEDLE_HF_REPO=<you>/<model>` and pass `--upload` to `needle build`; on any other machine `needle download <you>/<model>/tuned.cact` pulls it back down (or pass just `<you>/<model>` when the repo holds a single archive).
-
-Defaults: batch size 16, learning rate 0.0001 with warmup and cosine decay, gradient clipping at norm 1, rank 16, alpha 32, max length 1024, validation split 0.1, random seed 0. Use `--seed` to reproduce or intentionally vary LoRA initialization, validation selection, and epoch shuffling. The base checkpoint downloads from Hugging Face on first run.
-
-To grow a small hand written set, seed the generator with it (needs `OPENROUTER_API_KEY`; set `OPENROUTER_URL` to use another OpenAI compatible gateway):
-
-```sh
-needle generate-data --augment data.jsonl --num-samples 1000
-```
-
-The playground button labelled Finetune on these tools runs the same pipeline from the browser.
-
-Training is plain JAX, so it runs on any accelerator jax supports. On an NVIDIA machine install the CUDA build and the same command trains on the GPU, nothing else changes:
-
-```sh
-pip install "cactus-needle[train,gpu]"
-```
-
-Apple GPUs train through the jax metal plugin, which does not work past jax 0.4.38, so the `metal` extra pins an older stack:
-
-```sh
-pip install "cactus-needle[train,metal]"
-```
-
-Needle detects the Metal backend and adapts automatically: manual attention, no rematerialisation, unrolled layer stack, and the `ENABLE_PJRT_COMPATIBILITY` variable the plugin requires is set for you. Measured on an M5 Max: 0.71 seconds per step against 2.90 on CPU at the same shape, about 4 times faster, with a one time compile of about 23 seconds. Training runs in float32 on every backend.
-
-## Reading the loss
-
-The loss covers only the target: the reasoning line plus the JSON call. Much of the call is boilerplate the base model already predicts (the tool name, the braces, the field names), so training starts near 1.0 rather than near random. Judge a run by its trend, not its level.
-
-Step count is what small datasets get wrong. 200 examples at batch 16 is 13 steps per epoch, and the default 3 epochs is 39 steps total, which barely moves a rank 16 adapter at the default learning rate. For a few hundred examples run 10 to 30 epochs and expect a clear downward trend. If the curve sits at its starting value after a few hundred steps, raise the epochs first, then the learning rate.
-
-A validation loss prints at each epoch end (10 percent of examples are held out by default, `--val-split` to change). When it rises while the training loss keeps falling, the run is overfitting: stop there, or add data.
-
-## Sizing the dataset
-
-Tool selection moves first: a few hundred clean examples measurably improve which tool gets picked. Argument grounding moves later and needs more data, on the order of thousands of examples, with reasoning lines and varied phrasings and values. If evaluation shows correct tools with wrong argument values, the dataset is too small or too uniform, not mislabeled. For grounding heavy tasks `--lora-rank 32` doubles adapter capacity and the adapter stays tiny.
-
-For a large catalogue, consider two passes at inference instead of more training: one turn against the full catalogue to pick the tool, then one turn declaring only that tool, which constrains the grammar to exactly that call.
-
-## What finetuning does not change
-
-The confidence head. Scores are calibrated for the base model on its training mix and finetuning does not update the head, so the package disables them for tuned weights: `Needle(weights=...)` warns once at construction and reports `confidence` as None. Non English deployments of the base model should also treat the score with caution (correct Spanish calls have been measured at confidence 0.0).
-
-The tokenizer. Non English text fragments into roughly 1.7 times more tokens (measured on Spanish), which taxes both quality and the 256 token window.
-
-## Troubleshooting
-
-1. Loss goes NaN within the first steps on CPU: fixed, run `pip install --upgrade cactus-needle`.
-2. The tuned model answers "Sorry, I can't help with that" on everything: an old engine gated low confidence responses; the gate is removed as of engine 2.0.1. Upgrade the package.
-3. `failed to load weights`: the `.cact` format is tied to the engine version, so an archive exported by an older package will not load. Rebuild it with the current package version.
-4. Loss hovers at its starting value: see Reading the loss. The run is undertrained, not broken.
diff --git a/llms.txt b/llms.txt
index 4a7ac69..2396cbb 100644
--- a/llms.txt
+++ b/llms.txt
@@ -1,8 +1,8 @@
# Needle 3 (cactus-needle)
-> Needle 3 is a 25-121M parameter on-device model for tool calling, device control, and structured extraction. This Python package (`cactus-needle`) is inference + LoRA fine-tuning + export. Text goes in, a JSON tool call comes back; a byte-level grammar compiled from your schemas constrains every token, so the call is always well-formed. A sub-1 MB engine and the 35 MB `needle3.cact` weights are fetched once from Hugging Face and cached; nothing is compiled locally.
+> Needle 3 is a 29-121M parameter on-device model for tool calling, device control, and structured extraction. This Python package (`cactus-needle`) is inference + LoRA fine-tuning + export. Text goes in, a JSON tool call comes back; a byte-level grammar compiled from your schemas constrains every token, so the call is always well-formed. A sub-1 MB engine and the 35 MB `needle3.cact` weights are fetched once from Hugging Face and cached; nothing is compiled locally.
-This file is written for AI coding assistants. It is enough to write correct Needle code without reading the source. Copy the patterns; do not invent API that is not listed here.
+This file is written for AI coding assistants. It is enough to write correct Needle code without reading the source. Copy the patterns; do not invent API that is not listed here. The human guides are at https://cactuscompute.com/blog/needle-python-docs, https://cactuscompute.com/blog/designing-tools-for-needle, https://cactuscompute.com/blog/needle-confidence, https://cactuscompute.com/blog/structured-extraction-with-needle, https://cactuscompute.com/blog/finetuning-needle and https://cactuscompute.com/blog/needle-supported-devices.
## Install
@@ -23,7 +23,7 @@ pip install "cactus-needle[train]" # adds JAX + training deps for finetune
- `agent.run(query, max_steps=8, max_new_tokens=512) -> dict` - full agentic loop: model picks calls, Needle executes your Python functions, feeds results back, returns the final response with the executed tool results attached as `results`.
- `agent.complete(text, max_new_tokens=512) -> dict` - one turn; you execute the call and feed the result back yourself via the next `complete(...)`.
- `agent.reset()` - rewind the conversation, keep the tools loaded.
-- `needle.tool` - decorator that turns a function into a tool schema (attached as `fn._needle_tool`).
+- `needle.tool` - decorator that turns a function into a tool schema (attached as `fn._needle_tool`); `@needle.tool(triggers=[r"\b(turn|switch)\b.*\b(on|off)\b"])` adds case-insensitive regexes that route a matching request to this tool and require a call; a triggered call ships even below the confidence floor (raw schemas take the same `"triggers"` list beside `description`).
- `needle.Field(default=..., *, description, enum, const, ge, le, gt, lt, multiple_of, min_length, max_length, pattern, format, min_items, max_items, unique_items)` - per-argument constraints; attach inline with `typing.Annotated`.
- `needle.extract(text, schema, system=None, max_new_tokens=512, weights=None)` - one-shot extraction; returns a Pydantic instance if `schema` is a model, else a dict (or `None` if nothing matched). `weights` selects a tuned `.cact` and defaults to whatever the engine already has loaded, so extraction inherits the active tuned model unless you pass another path.
@@ -107,7 +107,7 @@ Every turn returns one dict:
- `Needle(system=...)` prefixes the local `date:` fact automatically unless the text already carries a date; pass `auto_date=False` to opt out.
- `validation.ungrounded`, when present, lists `tool.field` names whose value is not grounded in the input, including date arguments whose year matches nothing in the conversation or the `system` facts. `run()` refuses those calls with `{"error": "ungrounded field"}` unless `strict=False`.
- `extract()` declares the record as the only tool and takes the record from `function_calls` or, when the engine withheld it, from `suppressed_calls`; `strict` handles grounding.
-- Arguments arrive after the engine's deterministic repair step, which grounds them in the request text: a full name in one of a first/last pair is split, a truncated phone number is completed, quoted titles are restored verbatim, place queries recover dropped words and a following street address, weekday and `March 4th` style dates resolve against the `date:` fact, empty optional strings and duplicate calls are dropped. Enumerated parameters are never altered.
+- Arguments arrive after the engine's deterministic repair step, which grounds them in the request text: a full name in one of a first/last pair is split, a truncated phone number is completed, quoted titles are restored verbatim, place queries recover dropped words and a following street address, weekday and `March 4th` style dates resolve against the `date:` fact, polar enums and booleans follow the request verb (turn on, unlock, louder, turn down), an enum moves to the one option the request names, pickup and destination are read from `from X to Y` wording, a paired tool follows the request's polarity (`lock_door` to `unlock_door`), `by N` fills a tool's only number argument, an invented required slot is refilled from the request or its schema default, a required number with a default takes the default when no quantity is named, optional strings and numbers the conversation never states are dropped, empty optional strings and duplicate calls are dropped, and so is a call for something the request excludes or negates in its own clause.
## Behaviour contract (important for correct code)
@@ -132,13 +132,18 @@ if r["type"] == "call":
```python
r = agent.complete(user_text)
-calls = r.get("function_calls") or []
-if calls and r["confidence"] >= 0.8:
- execute(calls[0])
+calls = r["function_calls"]
+held = r["suppressed_calls"]
+if calls and r["confidence"] >= 0.7:
+ execute(calls) # sure: act
+elif calls or held:
+ confirm(calls or held, r["reasoning"]) # unsure: show the call, ask
else:
- confirm_or_reask()
+ say("I can't do that here") # nothing to do: refuse
```
+The engine already withholds calls below 0.1 into `suppressed_calls`; a tool with `triggers` always produces a call for a matching request, so the score decides whether to run it or confirm it.
+
## System facts (optional)
Pass environment state as facts, never instructions. Recognized keys: `date`, `locale`, `device`, `battery`, `network`, `location`, `user`, `assistant`.
@@ -184,12 +189,7 @@ Options: `finetune --lora-rank 16 --lora-alpha 32 --lr 1e-4 --batch-size 16 --ma
## Playground (browser)
-```sh
-needle playground # base model at http://127.0.0.1:7860
-needle playground --weights my.cact # a tuned model
-```
-
-Preset demos, editable tools/prompt, multi-turn follow-ups, and a "Finetune on these tools" button that runs the pipeline above and returns a downloadable `.cact`.
+Hosted at https://cactuscompute.com/needle: preset demos, editable tools/prompt, multi-turn follow-ups, and a "Finetune on these tools" button that runs the pipeline above and returns a downloadable `.cact`. `needle playground [--weights my.cact]` serves the same UI locally at http://127.0.0.1:7860. Local fine-tuning is 4-bit; the 2-bit post-training and quantisation behind the shipped model, enriched with Cactus proprietary datasets, run on the Cactus Platform at https://cactuscompute.com/dashboard.
## Environments (ready-made tool surfaces)
@@ -205,9 +205,9 @@ Adapt by swapping the Literal values for your own and keeping the shapes: closed
## CLI summary
- `needle run --checkpoint <base.safetensors or .pkl> --query "..." --tools tools.json` - JAX reference inference from a checkpoint (dev path; normal inference is the Python `Needle` API above).
-- `needle generate-data` / `needle finetune` / `needle build [checkpoint] [--lora adapter] [--layers N]` - the fine-tuning pipeline; fine-tuning runs at the full 20 layers, `build` merges the adapter, slices the `--layers` rung (default 20), exports at 4 bits, and fetches the published `needle3.cact` for its tokenizer on every run.
-- `needle playground` - browser UI.
-- `needle fetch [--platform-tag <tag>] [--out <dir>]` - pre-download the inference engine for this machine (or another platform) into the cache; prints the path. For air-gapped devices, also see `NEEDLE_LIB_PATH` and `HF_HUB_OFFLINE` in doc/apis.md.
+- `needle generate-data` / `needle finetune` / `needle build [checkpoint] [--lora adapter] [--layers N] [--platform <folder>] [--out <path>]` - the fine-tuning pipeline; fine-tuning runs at the full 20 layers, `build` merges the adapter, slices the `--layers` rung (default 20), exports at 4 bits, and fetches the published `needle3.cact` on every run (without an adapter at full depth it ships that published 2-bit archive as is). `--platform` also downloads that platform's engine and header and makes `--out` a folder holding them plus `needle3.cact`.
+- `needle playground` - the browser UI locally (hosted at https://cactuscompute.com/needle).
+- `needle fetch [--platform-tag <tag>] [--out <dir>]` - pre-download the inference engine for this machine (or another platform) into the cache; prints the path. For air-gapped devices, also see `NEEDLE3_LIB_PATH` and `HF_HUB_OFFLINE` at https://cactuscompute.com/blog/needle-python-docs.
- `needle download needle3 | needle3.safetensors | <org>/<repo>[/<file>.cact] | <platform> [--out <dir>]` - pull the base Needle 3 archive, a checkpoint to fine-tune, a published `.cact` (single-archive repos need only `<org>/<repo>`), or pass a platform folder name (`macos-arm64`, `linux-x86_64`, `wasm`, `wasm-component`, ...) to fetch that platform's engine files into `<out>/<platform>/`. Native runners are marked executable; `wasm-component` contains the WASI Preview 2 component and WIT contract.
## Common mistakes to avoid
diff --git a/needle/agent/fetch.py b/needle/agent/fetch.py
index 5985c1b..41a35ed 100644
--- a/needle/agent/fetch.py
+++ b/needle/agent/fetch.py
@@ -93,7 +93,7 @@ def _register_download(generation=2):
pass
-def download_platform(name, out_dir, generation=2):
+def download_platform(name, out_dir, generation=2, dest=None):
import shutil
import stat
from huggingface_hub import hf_hub_download, list_repo_files
@@ -101,7 +101,9 @@ def download_platform(name, out_dir, generation=2):
repo = engine_repo(generation)
_register_download(generation)
files = [f for f in list_repo_files(repo) if f.startswith(name + "/")]
- dest = os.path.join(out_dir, name)
+ if name not in PLATFORMS or not files:
+ raise FileNotFoundError(f"{name} is not a published platform folder in {repo}")
+ dest = dest or os.path.join(out_dir, name)
os.makedirs(dest, exist_ok=True)
out = []
for f in files:
diff --git a/needle/agent/tools.py b/needle/agent/tools.py
index e16b7e6..c4d9bcc 100644
--- a/needle/agent/tools.py
+++ b/needle/agent/tools.py
@@ -170,6 +170,11 @@ def pydantic_schema(model: type) -> dict:
return out
-def tool(fn: Callable) -> Callable:
- fn._needle_tool = build_schema(fn)
- return fn
+def tool(fn: Callable | None = None, *, triggers: list[str] | None = None) -> Callable:
+ def attach(target: Callable) -> Callable:
+ schema = build_schema(target)
+ if triggers:
+ schema["triggers"] = list(triggers)
+ target._needle_tool = schema
+ return target
+ return attach(fn) if fn is not None else attach
diff --git a/needle/cli.py b/needle/cli.py
index 9b59d17..191b5b7 100644
--- a/needle/cli.py
+++ b/needle/cli.py
@@ -136,6 +136,8 @@ def main():
print(HELP)
sys.exit(0)
+ from .agent import fetch
+
parser = argparse.ArgumentParser(prog="needle", add_help=False)
sub = parser.add_subparsers(dest="command")
p = sub.add_parser("run")
@@ -190,6 +192,9 @@ def main():
p.add_argument("--upload", action="store_true", help="Push the .cact to $NEEDLE_HF_REPO")
p.add_argument("--layers", type=int, default=None,
help="Export the N-layer rung of the base (2..20); default the full 20")
+ p.add_argument("--platform", type=str, default=None, choices=fetch.PLATFORMS,
+ help="Also download that platform's engine and header, and place the "
+ "archive beside them as needle3.cact (--out is then a directory)")
p = sub.add_parser("download")
p.add_argument("spec", type=str,
@@ -198,16 +203,16 @@ def main():
"folder (e.g. macos-arm64), or a Hugging Face spec: "
"<org>/<repo>/<file>.cact, or <org>/<repo> if it holds one archive")
p.add_argument("--out", type=str, default=".", help="Directory to place the files")
- p.add_argument("--generation", type=int, choices=[2, 3], default=2,
- help="Engine generation when downloading a platform build (default: 2)")
+ p.add_argument("--generation", type=int, choices=[2, 3], default=3,
+ help="Engine generation when downloading a platform build (default: 3)")
p = sub.add_parser("fetch")
p.add_argument("--out", type=str, default=None,
help="Directory to place the engine (default: the cache)")
p.add_argument("--platform-tag", type=str, default=None,
help="Fetch the build for another device, e.g. manylinux2014_aarch64")
- p.add_argument("--generation", type=int, choices=[2, 3], default=2,
- help="Needle engine generation to fetch (default: 2)")
+ p.add_argument("--generation", type=int, choices=[2, 3], default=3,
+ help="Needle engine generation to fetch (default: 3)")
p = sub.add_parser("playground")
p.add_argument("--weights", type=str, default=None,
diff --git a/needle/model/export.py b/needle/model/export.py
index cb0c99f..faa2f8f 100644
--- a/needle/model/export.py
+++ b/needle/model/export.py
@@ -366,6 +366,14 @@ def _tokenizer_pieces(blob):
return struct.unpack_from(_TK_HDR, blob, 0)[0]
+def read_layers(path):
+ with open(path, "rb") as f:
+ hdr = struct.unpack(_HDR_FMT, f.read(struct.calcsize(_HDR_FMT)))
+ if hdr[0] != TAG:
+ raise ValueError(f"{path} is not a Needle 3 .cact archive")
+ return hdr[10]
+
+
def read_tokenizer_blob(path):
"""The packaged tokenizer of a .cact archive, read without unpacking its weights."""
with open(path, "rb") as f:
diff --git a/needle/model/finetune.py b/needle/model/finetune.py
index eddab9c..a246170 100644
--- a/needle/model/finetune.py
+++ b/needle/model/finetune.py
@@ -431,35 +431,56 @@ def finetune_local(args, progress=None):
def build_main(args):
+ import shutil
import jax.numpy as jnp
from ..agent import fetch
from .run import load_checkpoint
from .architecture import effective_kv_window
- from .export import read_tokenizer_blob, write_export
+ from .export import read_layers, read_tokenizer_blob, write_export
from .quantize import WEIGHT_BITS
base_archive = fetch.fetch_weights(3, force=True)
print(f" {'base':<9} {base_archive} {os.path.getsize(base_archive) / 1e6:.2f} MB")
- adapter = read_adapter(args.lora) if args.lora else None
- checkpoint = args.checkpoint
- if not checkpoint and adapter and adapter.get("base") and os.path.exists(adapter["base"]):
- checkpoint = adapter["base"]
- checkpoint = checkpoint or DEFAULT_BASE
- params, config, _ = load_checkpoint(checkpoint, return_run=True)
- if args.lora:
- lora = {tuple(key.split("/")): {"A": jnp.asarray(v["A"]), "B": jnp.asarray(v["B"])}
- for key, v in adapter["lora"].items()}
- params = merge_lora(params, lora, adapter["scale"])
- print(f" {'merged':<9} {len(lora)} weight groups {args.lora}")
- params, config = rung(params, config, getattr(args, "layers", None))
- print(f" {'depth':<9} {config.num_layers} layers")
-
- out = args.out or (os.path.splitext(os.path.basename(checkpoint))[0] + ".cact")
- info = write_export(params, config, out, bits=WEIGHT_BITS,
- tokenizer=read_tokenizer_blob(base_archive),
- kv_window=effective_kv_window(config))
- print(f" {'wrote':<9} {info['path']} {info['bytes'] / 1e6:.2f} MB {info['tensors']} tensors W{WEIGHT_BITS}A8")
- print(f" {'next':<9} needle.Needle(weights={out!r}, tools=[...])")
+ platform = getattr(args, "platform", None)
+ layers = getattr(args, "layers", None)
+ folder = None
+ if platform:
+ folder = os.path.abspath(args.out or platform)
+ for path in fetch.download_platform(platform, os.path.dirname(folder), generation=3, dest=folder):
+ print(f" {'engine':<9} {path} {os.path.getsize(path) / 1e6:.2f} MB")
+ out = os.path.join(folder, fetch.base_weights(3))
+ else:
+ out = args.out
+ if not args.lora and (not layers or layers == read_layers(base_archive)):
+ if not out:
+ raise SystemExit("pass --out <archive.cact>, or --platform to build a runnable folder")
+ shutil.copyfile(base_archive, out)
+ print(f" {'wrote':<9} {out} {os.path.getsize(out) / 1e6:.2f} MB the published base archive")
+ else:
+ adapter = read_adapter(args.lora) if args.lora else None
+ checkpoint = args.checkpoint
+ if not checkpoint and adapter and adapter.get("base") and os.path.exists(adapter["base"]):
+ checkpoint = adapter["base"]
+ checkpoint = checkpoint or DEFAULT_BASE
+ params, config, _ = load_checkpoint(checkpoint, return_run=True)
+ if args.lora:
+ lora = {tuple(key.split("/")): {"A": jnp.asarray(v["A"]), "B": jnp.asarray(v["B"])}
+ for key, v in adapter["lora"].items()}
+ params = merge_lora(params, lora, adapter["scale"])
+ print(f" {'merged':<9} {len(lora)} weight groups {args.lora}")
+ params, config = rung(params, config, layers)
+ print(f" {'depth':<9} {config.num_layers} layers")
+ out = out or (os.path.splitext(os.path.basename(checkpoint))[0] + ".cact")
+ info = write_export(params, config, out, bits=WEIGHT_BITS,
+ tokenizer=read_tokenizer_blob(base_archive),
+ kv_window=effective_kv_window(config))
+ print(f" {'wrote':<9} {info['path']} {info['bytes'] / 1e6:.2f} MB {info['tensors']} tensors W{WEIGHT_BITS}A8")
+ if folder:
+ runner = next((n for n in ("needle", "needle.exe") if os.path.exists(os.path.join(folder, n))), None)
+ if runner:
+ print(f" {'next':<9} {os.path.join(folder, runner)} --model {os.path.basename(out)} --tools tools.json --serve")
+ else:
+ print(f" {'next':<9} needle.Needle(weights={out!r}, tools=[...])")
if args.upload:
repo = os.environ.get("NEEDLE_HF_REPO")
diff --git a/tests/test_build.py b/tests/test_build.py
index 338b2c0..19968da 100644
--- a/tests/test_build.py
+++ b/tests/test_build.py
@@ -6,8 +6,9 @@ import pytest
pytestmark = pytest.mark.slow
-def _build_args(checkpoint, out, lora=None):
- return types.SimpleNamespace(checkpoint=checkpoint, lora=lora, out=out, upload=False)
+def _build_args(checkpoint, out, lora=None, layers=None, platform=None):
+ return types.SimpleNamespace(checkpoint=checkpoint, lora=lora, out=out, upload=False,
+ layers=layers, platform=platform)
def test_build_exports_loadable_cact(tiny_checkpoint, tmp_path, published_base):
@@ -92,3 +93,25 @@ def test_build_layers_exports_the_rung(tiny_checkpoint, tmp_path, published_base
build_main(args)
header, _ = read_export(out)
assert header["num_layers"] == 3
+
+
+def test_build_platform_places_the_engine_and_the_archive_together(tiny_checkpoint, tmp_path, published_base, monkeypatch):
+ from needle.agent import fetch
+ from needle.model.finetune import build_main
+ from needle.model.export import read_layers
+
+ def fake_platform(name, out_dir, generation=3, dest=None):
+ os.makedirs(dest, exist_ok=True)
+ runner = os.path.join(dest, "needle")
+ with open(runner, "wb") as handle:
+ handle.write(b"engine")
+ return [runner]
+
+ monkeypatch.setattr(fetch, "download_platform", fake_platform)
+ folder = str(tmp_path / "pi")
+ build_main(_build_args(None, folder, platform="linux-arm64"))
+ assert os.path.exists(os.path.join(folder, "needle"))
+ archive = os.path.join(folder, "needle3.cact")
+ assert read_layers(archive) == read_layers(fetch.fetch_weights(3))
+ build_main(_build_args(tiny_checkpoint, folder, platform="linux-arm64", layers=3))
+ assert read_layers(archive) == 3