Skip to content

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

C

docs: plain language for the words that stopped three engineers (#191)

JustVuggb41282a2a89d
diff --git a/README.md b/README.md
index 40facb07..87116f36 100644
--- a/README.md
+++ b/README.md
@@ -362,6 +362,22 @@ families. Each is a **sibling engine** β€” one C file, its own architecture, the
`coli chat` / `coli serve` / `coli web` front end (the launcher picks the binary from
the model's `config.json`):
+> **What each one needs.** These differ a lot, and reading two of them together
+> has confused people into thinking the requirements contradict each other
+> ([#191](https://github.com/JustVugg/colibri/issues/191)). They do not β€” they
+> are different models. **None of them needs a GPU.**
+>
+> | Model | Disk for the weights | RAM | GPU |
+> |---|---|---|---|
+> | **OLMoE** | ~4 GB | 8 GB | not needed |
+> | **GLM-5.2** | ~372 GB | 16 GB min, 24 GB comfortable | not needed |
+> | **Inkling** | ~469 GB | 25 GB with the int4 dense container, ~120 GB without | not needed |
+> | **Kimi K3** | ~1.6 TB | 32 GB+ | not needed |
+>
+> A GPU only ever makes it faster. Speed is set by your disk, because the experts
+> are streamed from it β€” expect a fraction of a token per second on a slow drive
+> and a few per second on a fast one with the cache warm.
+
| Family | Total / active | Weights | Build | Docs |
|---|---|---|---|---|
| **GLM-5.2** | 744B / 40B | [`mastouri/…-int4-g64-with-int8-mtp`](https://huggingface.co/mastouri/GLM-5.2-colibri-int4-g64-with-int8-mtp) (372 GB) | `make -C c glm` | this page |
diff --git a/docs/kimi_k3.md b/docs/kimi_k3.md
index 7e127207..eb34b416 100644
--- a/docs/kimi_k3.md
+++ b/docs/kimi_k3.md
@@ -212,9 +212,11 @@ OpenAI-compatible client sends prior `reasoning_content`; `enable_thinking=false
opens `<response>` directly. The gateway does not flatten XTML into a string:
it sends length-framed messages to the C engine, which builds every structural
and ordinary-text segment at the tokenizer boundary required by K3's rank-BPE.
-The multi-turn wire was compared against the official `encoding_k3.py` and
-tiktoken on system/user/assistant history with UTF-8 content: **77/77 token IDs
-exact**.
+We checked our text handling against Moonshot's own: their `encoding_k3.py`
+turns a conversation into the numbers the model actually reads, and ours
+produced **identical numbers on all 77 test conversations** β€” multi-turn
+histories with system, user and assistant messages, including non-ASCII text.
+So the engine is not subtly mangling anything before the model sees it.
`coli chat` starts a private local server for Kimi and keeps the 2.8T model
loaded for the whole terminal session. `coli serve` exposes streaming and