# Colibrì

A pure C inference engine that runs very large mixture-of-experts language models on consumer hardware by treating disk, RAM and VRAM as one memory tier.

Colibrì is an inference engine for running large mixture-of-experts language models, from 744 billion up to 2.8 trillion parameters, on consumer and mixed hardware. It is written in pure C with no external engine dependencies, and streams model experts from disk, treating storage, RAM and VRAM as a single tiered memory hierarchy rather than requiring the whole model to fit in fast memory at once.

Nine model families are supported, including GLM, Kimi K3, DeepSeek V4 Flash and Qwen variants, each implemented in a single C file sharing the same chat, serve and web front ends. The project describes itself as both a usable inference engine and an open research platform for pursuing inference-side performance across model formats, memory hierarchy, storage I/O, scheduling and CPU/GPU overlap, with a stated policy that experiments must be reproducible and must never silently change model precision or routing behavior, even if that costs speed.

Website: https://justvugg.github.io/colibri