What is this?
Colibri is an open-source engine for running frontier mixture-of-experts (MoE) language models on hardware you already own — written in pure C with zero dependencies.
Colibri is an open-source engine for running frontier mixture-of-experts (MoE) language models on hardware you already own — written in pure C with zero dependencies.So is it a library I import, or a program I run?It is a compact inference engine you run: the project pitches itself as a tiny engine for executing immense MoE models locally, not a cloud service or an SDK.
It is a compact inference engine you run: the project pitches itself as a tiny engine for executing immense MoE models locally, not a cloud service or an SDK.How do I get it running?The repo builds releases in CI, so the practical first step is a prebuilt binary from GitHub Releases — or clone the repo and compile the zero-dependency C source yourself.
The repo builds releases in CI, so the practical first step is a prebuilt binary from GitHub Releases — or clone the repo and compile the zero-dependency C source yourself.Do I need CUDA, Python or other runtimes installed first?No — the engine is pure C with zero dependencies, so there is no framework, runtime or library to install before building or running it.
No — the engine is pure C with zero dependencies, so there is no framework, runtime or library to install before building or running it.Does everything run locally on my machine?Yes — the whole premise is executing models on hardware you already own, so inference happens locally rather than through a cloud API.
Yes — the whole premise is executing models on hardware you already own, so inference happens locally rather than through a cloud API.Do I need a big GPU to run a frontier-size model?No — instead of holding all weights in memory, Colibri streams each expert from disk as needed, so model size is decoupled from your GPU and RAM budget.
No — instead of holding all weights in memory, Colibri streams each expert from disk as needed, so model size is decoupled from your GPU and RAM budget.What models can it run?It is built for frontier-scale MoE transformer models — the sparse architecture where each token activates only a subset of experts, which is exactly what makes disk-streaming viable.
It is built for frontier-scale MoE transformer models — the sparse architecture where each token activates only a subset of experts, which is exactly what makes disk-streaming viable.Why is pure C with zero dependencies a selling point?Nothing to install beyond a C compiler: the engine stays tiny and portable, matching the project motto 'Tiny engine, immense model'.
Nothing to install beyond a C compiler: the engine stays tiny and portable, matching the project motto 'Tiny engine, immense model'.Is Colibri free to use commercially?Yes — the repository is Apache-2.0 licensed, so the engine is free to use, modify and ship, including commercially, as long as the license terms are followed.
Yes — the repository is Apache-2.0 licensed, so the engine is free to use, modify and ship, including commercially, as long as the license terms are followed.How active is the project?Very: over 41,000 stars, and the repo ships CI workflows for checks, releases and a documentation site — though the project itself is young, started in July 2026.
Very: over 41,000 stars, and the repo ships CI workflows for checks, releases and a documentation site — though the project itself is young, started in July 2026.What are similar projects?warp, kimi-k3-in-c, Strata, slotstream, airllm, FreeToken, Edge0, StartLux-Decision
warp, kimi-k3-in-c, Strata, slotstream, airllm, FreeToken, Edge0, StartLux-DecisionWhich topics is it in?on-device-ai, on-device, inference-engine, mixture-of-experts, moe, llm-inference, transformer, local-llm, c, inference, local-first, deep-learning
on-device-ai, on-device, inference-engine, mixture-of-experts, moe, llm-inference, transformer, local-llm, c, inference, local-first, deep-learning