Skip to content

Run frontier MoE models on hardware you already own — pure C, zero deps, experts streamed from disk. Tiny engine, immense model. 🐦

C

docs(readme): the sixth family, and one logo across every translation

JustVuggc4e5b69c9a14
diff --git a/README.it.md b/README.it.md
index ffc5f6b2..5d61892c 100644
--- a/README.it.md
+++ b/README.it.md
@@ -1,5 +1,5 @@
<p align="center">
- <img src="assets/colibri.svg" width="500" alt="colibrì — motore piccolo, modello immenso">
+ <img src="assets/colibri-logo.svg" width="560" alt="colibrì — motore piccolo, modello immenso">
</p>
<p align="center">
@@ -12,9 +12,9 @@ miliardi a 2,8 mila miliardi di parametri** — su hardware consumer ed eterogen
in C puro e senza dipendenze del motore, trattando storage, RAM e VRAM come
un'unica gerarchia di inferenza.
-Oggi girano quattro famiglie: **GLM-5.2** (744B), **Inkling** (975B), **Kimi K3**
-(2,8T) e **OLMoE** (7B) — un file C ciascuna, la stessa interfaccia `coli chat` /
-`coli serve` / `coli web`. [Elenco completo](README.md#other-supported-models)
+Oggi girano sei famiglie: **GLM-5.2** (744B), **Inkling** (975B), **Kimi K3**
+(2,8T), **DeepSeek V4 Flash** (284B), **Qwen3.6** (35B-A3B) e **OLMoE** (7B) — un
+file C ciascuna, la stessa interfaccia `coli chat` / `coli serve` / `coli web`. [Elenco completo](README.md#other-supported-models)
> **Colibrì è un motore di inferenza che puoi usare oggi, e una piattaforma di
> ricerca aperta.** Il suo obiettivo principale è migliorare le prestazioni di
diff --git a/README.md b/README.md
index 384ec3b6..164a2aae 100644
--- a/README.md
+++ b/README.md
@@ -18,9 +18,9 @@ parameters** — on consumer and heterogeneous hardware, in pure C with zero
engine dependencies, by treating storage, RAM, and VRAM as a single inference
hierarchy (AI memory multitiering).
-Five families run today: **GLM-5.2** (744B), **Inkling** (975B), **Kimi K3**
-(2.8T), **DeepSeek V4 Flash** (284B) and **OLMoE** (7B) — one C file each, the
-same `coli chat` / `coli serve` / `coli web` front end.
+Six families run today: **GLM-5.2** (744B), **Inkling** (975B), **Kimi K3**
+(2.8T), **DeepSeek V4 Flash** (284B), **Qwen3.6** (35B-A3B) and **OLMoE** (7B) —
+one C file each, the same `coli chat` / `coli serve` / `coli web` front end.
[Full roster ↓](#other-supported-models)
> **Colibrì is an inference engine you can run today, and an open research
@@ -390,7 +390,7 @@ the full 756 GB on disk at once:
#### Other supported models
-GLM-5.2 is the reference model, but the same streaming approach runs three more
+GLM-5.2 is the reference model, but the same streaming approach runs five more
families. Each is a **sibling engine** — one C file, its own architecture, the same
`coli chat` / `coli serve` / `coli web` front end (the launcher picks the binary from
the model's `config.json`):
@@ -407,6 +407,7 @@ the model's `config.json`):
> | **Inkling** | ~469 GB | 25 GB with the int4 dense container, ~120 GB without | not needed |
> | **Kimi K3** | ~1.6 TB | 32 GB+ | not needed |
> | **DeepSeek V4 Flash** | ~167 GB | 16 GB min, 32 GB comfortable | optional; an NVIDIA card (any sm_80+, best on RTX 50) makes prefill 5-10x and decode ~2.5x faster |
+> | **Qwen3.6-35B-A3B** | ~20 GB (int4-gs64 container) | 24 GB (needs full RAM residency) | optional; the CUDA VRAM expert tier measured **1.44 -> 10.05 tok/s (7.0x)** on two 8 GB cards, output bit-identical to CPU |
>
> A GPU only ever makes it faster. Speed is set by your disk, because the experts
> are streamed from it — expect a fraction of a token per second on a slow drive
@@ -418,8 +419,17 @@ the model's `config.json`):
| **Inkling** (Thinking Machines) | 975B / 41B | [`nbeerbower/Inkling-colibri-int4`](https://huggingface.co/nbeerbower/Inkling-colibri-int4) (469 GB) | `make -C c inkling` | [inkling.md](docs/inkling.md) |
| **Kimi K3** (Moonshot) | 2.8T / 104B | [`moonshotai/Kimi-K3`](https://huggingface.co/moonshotai/Kimi-K3) — original checkpoint, routed experts stay **native MXFP4** | `make -C c kimi_k3` | [kimi_k3.md](docs/kimi_k3.md) |
| **DeepSeek V4 Flash** | 284B / 13B | official sharded checkpoint — routed experts stay **native fp4**, dense stays fp8-e4m3 | `make -C c deepseek-v4` | [deepseek-v4.md](docs/deepseek-v4.md) |
+| **Qwen3.6** (Alibaba) | 35B / 3B | [`Kreuzzelg/qwen36-35b-a3b-colibri-i4-gs64`](https://huggingface.co/Kreuzzelg/qwen36-35b-a3b-colibri-i4-gs64) (~20 GB, **recommended**) — hybrid Gated Attention + Gated DeltaNet | `make -C c qwen36` (`CUDA=1` for the VRAM expert tier) | [qwen36.md](docs/qwen36.md) |
| **OLMoE** (AI2) | 7B / 1B | converted with `c/tools/convert_olmoe_merged.py` — **int8** container, ~7 GB | `make -C c olmoe` | — |
+Qwen3.6 ships three pre-converted containers: **int4-gs64** (recommended — measured
+cosine to the int8 anchor 0.98777 → 0.99313 and KL 0.109 → 0.080 against per-row, i.e.
+~44% less quantization error), [int4 per-row](https://huggingface.co/Kreuzzelg/qwen36-35b-a3b-colibri-i4)
+as the A/B baseline, and [KAT-Coder v2.5](https://huggingface.co/Kreuzzelg/kat-coder-v2.5-dev-colibri-i4-gs64),
+which the same engine runs unchanged — any architecture-identical checkpoint works
+without a code path of its own. With `CUDA=1` the VRAM expert tier measured
+**1.44 → 10.05 tok/s (7.0×) on two 8 GB cards**, output bit-identical to the CPU path.
+
Kimi K3 needs no conversion: its QAT-trained MXFP4 experts are streamed straight from
the original Hugging Face shards, and the bf16 dense set is quantized at load time.
Inkling ships int4 experts but **bf16 dense weights** (49.4 GB resident); on a host
diff --git a/README.zh-CN.md b/README.zh-CN.md
index 4274eb23..c8341105 100644
--- a/README.zh-CN.md
+++ b/README.zh-CN.md
@@ -1,5 +1,5 @@
<p align="center">
- <img src="assets/colibri.svg" width="500" alt="colibrì——小巧引擎,庞大模型">
+ <img src="assets/colibri-logo.svg" width="560" alt="colibrì——小巧引擎,庞大模型">
</p>
<p align="center">
@@ -10,9 +10,9 @@
**小巧引擎,庞大模型。**在消费级与异构硬件上运行**前沿 MoE 模型——从 744B 到
2.8T 参数**——以引擎零依赖的纯 C 实现,将存储、RAM 与 VRAM 视为统一的推理层级。
-目前可运行四个模型家族:**GLM-5.2**(744B)、**Inkling**(975B)、**Kimi K3**
-(2.8T)与 **OLMoE**(7B)——各自一个 C 文件,共用同一套 `coli chat` /
-`coli serve` / `coli web` 前端。[完整列表](README.md#other-supported-models)
+目前可运行六个模型家族:**GLM-5.2**(744B)、**Inkling**(975B)、**Kimi K3**
+(2.8T)、**DeepSeek V4 Flash**(284B)、**Qwen3.6**(35B-A3B)与 **OLMoE**(7B)
+——各自一个 C 文件,共用同一套 `coli chat` / `coli serve` / `coli web` 前端。[完整列表](README.md#other-supported-models)
> **Colibrì 既是今天就能运行的推理引擎,也是一个开放的研究平台。**它的首要目标是在
> 完整的软硬件边界上追求推理侧性能——模型格式、内存层级、存储 I/O、放置、调度、内核、
diff --git a/README.zh-TW.md b/README.zh-TW.md
index 2809dd19..2cb90082 100644
--- a/README.zh-TW.md
+++ b/README.zh-TW.md
@@ -1,5 +1,5 @@
<p align="center">
- <img src="assets/colibri.svg" width="500" alt="colibrì——小巧引擎,龐大模型">
+ <img src="assets/colibri-logo.svg" width="560" alt="colibrì——小巧引擎,龐大模型">
</p>
<p align="center">
@@ -10,9 +10,9 @@
**小巧引擎,龐大模型。**在消費級與異質硬體上執行**前沿 MoE 模型——從 744B 到
2.8T 參數**——以引擎零相依套件的純 C 實作,將儲存、RAM 與 VRAM 視為統一的推論階層。
-目前可執行四個模型家族:**GLM-5.2**(744B)、**Inkling**(975B)、**Kimi K3**
-(2.8T)與 **OLMoE**(7B)——各自一個 C 檔案,共用同一套 `coli chat` /
-`coli serve` / `coli web` 前端。[完整清單](README.md#other-supported-models)
+目前可執行六個模型家族:**GLM-5.2**(744B)、**Inkling**(975B)、**Kimi K3**
+(2.8T)、**DeepSeek V4 Flash**(284B)、**Qwen3.6**(35B-A3B)與 **OLMoE**(7B)
+——各自一個 C 檔案,共用同一套 `coli chat` / `coli serve` / `coli web` 前端。[完整清單](README.md#other-supported-models)
> **Colibrì 既是今天就能執行的推論引擎,也是一個開放的研究平台。**它的首要目標是在
> 完整的軟硬體邊界上追求推論側效能——模型格式、記憶體階層、儲存 I/O、配置、排程、核心、