Colibrì

Colibrì runs frontier mixture-of-experts models on consumer and heterogeneous hardware by treating VRAM, RAM and storage as a single memory hierarchy rather than assuming a model must fit in VRAM. Nine families run today — GLM-5.2/5.3 at 744B, GLM-5.3-Flash at 321B with vision, Inkling at 975B, Kimi K3 at 2.8T, DeepSeek V4 Flash at 284B and V4.1 Flash at 552B, Qwen3.8-Flash-Next at 125B, Qwen3.6 at 35B-A3B and OLMoE at 7B — each as one C file behind the same coli chat, coli serve and coli web front end.

Who built itThe repository accumulated 2,803 commits between July and October 2026, and the two largest shares belong to Vincenzo Fornaro (837) and the GitHub account JustVugg (593). Behind them the history is unusually wide for a project this young — ZacharyZcR 216, monotophic 141, Vincenzo 63, Steve Markgraf 60, woolcoxm 55 — in a contributor list that runs to well over a hundred names.

1. 01

   A whole engine in three months, and who wrote it

   The repository was created on 2026-07-01 and has been pushed to every day since; 2,803 commits landed in July (1,028), August (985) and September (790), across 857 files. The authorship is the part a reader should look at first: 591 commits carry a co-author trailer and roughly 575 of them name a Claude model — Fable 5 on 166, Opus 5 on 153, Opus 4.8 on 85, Opus 5 (1M context) on 57, Fable 5.1 on 41, Opus 4.8 (1M context) on 20, and smaller counts for Opus 5.5, Sonnet 4.6 and Sonnet 5, with Claude Code, Claude and Cursor on a handful more. Sixteen trailers name people, most of them the maintainers. The commits are not only attributed at the trailer level either: several appear under machine-generated author names of the form codex 20260923-120015-1331522267.

2. 02

   The claim is about where the model lives, not how fast it runs

   The README states the design in two sentences: frontier MoE models of 744B to 2.8T parameters run on consumer hardware by treating storage, RAM and VRAM as one inference hierarchy, in pure C with zero engine dependencies. The project calls this AI memory multitiering and frames itself as a research platform first. The commitment that follows from that framing is stated as a rule rather than a hope: there is no SLA on speed, and a hard guarantee on semantics — experiments have to earn their place through reproducible end-to-end measurements, and the default policy never silently changes model precision or router semantics. Insufficient fast memory may cost speed; it may not quietly redefine the model.

3. 03

   Nine model families, one C file each

   What runs today is listed in the README with parameter counts: GLM-5.2/5.3 at 744B, GLM-5.3-Flash at 321B with vision, Inkling at 975B, Kimi K3 at 2.8T, DeepSeek V4 Flash at 284B, DeepSeek V4.1 Flash at 552B with vision, Qwen3.8-Flash-Next at 125B plus a 51B n-gram, Qwen3.6 at 35B-A3B, and OLMoE at 7B. Each family is one C file, and all of them sit behind the same three front ends, so the model is a configuration of the engine rather than a fork of it. The documentation follows the same split: docs/FORMATS.md at 38 KB, docs/ENVIRONMENT.md at 61 KB, docs/metal_implementation.md at 47 KB, docs/windows.md at 35 KB, docs/benchmarks.md at 28 KB, and one document per model family — qwen38.md at 29 KB, deepseek-v41.md at 26 KB, deepseek-v4.md at 25 KB.

4. 04

   Tests are the largest thing in the repository

   Of the 857 files, 390 sit in c/tests and account for 3.6 MB against 6.3 MB of engine code in c/ — a ratio that follows from the semantics guarantee, since a rule about never silently changing precision is only as good as the checks that would notice. The contributor list is also unusually broad for a project three months old: alongside the two maintainers there are more than a hundred named contributors, from twenty- and sixty-commit contributors down to a long tail of single commits.