Pure C · zero dependencies · Apache 2.0

Run models bigger than your machine

colibrì streams Mixture-of-Experts weights from disk instead of loading them, so a 2.8-trillion-parameter model runs on a 32 GB box. Six model families, one engine, CPU or GPU.

25,157
GitHub stars
2,730
forks
100+
contributors
510
pull requests merged
6
families · 4 backends · 3 platforms
All of it in six weeks, in the open: issues filed from strangers' machines, reproduced and fixed the same day. Currently shipping v1.10.1.

Frontier models should not be sealed inside datacenters.

Anyone curious enough should be able to open one up: run it, watch every expert fire, and make it better.

Read the engine →
How it works

Weights are not state to hold. They are data to stage.

A token through GLM-5.2 activates only ~40B parameters, and just ~11 GB of those change from token to token: the routed experts. colibrì keeps the dense weights resident and treats the 19,456 experts as a storage hierarchy. Measured routing heat decides which experts earn VRAM, which earn pinned RAM, and which stream from NVMe, and the router runs a layer ahead so prefetch hides the latency.

Everything is validated token-exact against the reference transformers implementation. Speed never buys drift.

The models

Eight families run today, one engine each.

GLM-5.2

The flagship target: an int4 container, token-exact versus reference, with the full expert atlas published.

reasoningtools
MakerZ.ai
Size · RAM744B MoE · from 16 GB
Run it →

GLM-5.3-Flash

The flagship’s fast sibling: the same 40B active at less than half the total size, and it takes images through the same server.

reasoningtoolsvision
MakerZ.ai
Size · RAM321B MoE · 25 GB
Run it →

Qwen3.8-Flash-Next

The first family to arrive complete: tool calling and a vision tower verified against the upstream reference, on day one.

reasoningtoolsvision
MakerAlibaba Qwen
Size · RAM125B+51B MoE · from 16 GB
Run it →

Kimi K3

The largest family we run: gated recurrence plus MLA attention, 2.8T parameters, now with a Metal backend on Apple Silicon.

reasoningtools
MakerMoonshot AI
Size · RAM2.8T MoE · from 32 GB
Run it →

Inkling

A 975B reasoning MoE with an audio tower. Mixed-precision staging fits it on a consumer box.

reasoningaudio
MakerThinking Machines
Size · RAM975B MoE · 25 GB
Run it →

DeepSeek V4 Flash

Sparse attention, multi-token prediction and a trained drafter: the most instrumented engine here.

reasoningtools
MakerDeepSeek AI
Size · RAM284B MoE · from 16 GB
Run it →

Qwen3.6

Gated attention plus Gated DeltaNet. Its CUDA tier keeps hot experts in VRAM: 7x faster than CPU on two 8 GB cards, output bit-identical.

reasoning
MakerAlibaba Qwen
Size · RAM35B MoE · 24 GB
Run it →

OLMoE

The small research workhorse. Quantization A/Bs and quality ablations run here first.

MakerAllen AI
Size · RAM7B MoE · 8 GB
Run it →

The mind, measured.

Every point behind this text is a real measured expert: 13,260 characterised experts, specialists clustering by topic, generalists drifting to the centre. The first expert atlas of a 700B-class model, mapped on ordinary machines.

drag to spin · position is measured routing affinity, not a learned embedding
Hardware

Same engine, any machine.

Measured decode speed on real community machines. The hardware only changes where the experts live, not what the model answers.

…and 15+ more community machines, from an M4 Pro Mac Mini to a 430 GB EPYC server. Full tables, methodology and quality ablations: docs/benchmarks.md

Come build the microscope.

bash
# one C file, no dependencies
$ git clone https://github.com/JustVugg/colibri
$ make -C colibri/c

# point it at a checkpoint and talk to it
$ ./coli chat --model /path/to/glm-5.2

# or serve an OpenAI-compatible API + dashboard
$ ./coli web --model /path/to/glm-5.2