colibri Hits GitHub Trending: Pure-C Engine Runs a 744B Model on Ordinary PCs

What happened

colibri was created on GitHub on July 1, 2026, has shipped 20 releases and is licensed under Apache-2.0. Its stated idea: a MoE model is huge on disk but uses only a small part of itself per token. For GLM-5.2, with 744B parameters in total, about 40B are active per token, and the routed experts that change from token to token are about 11 GB. colibri therefore keeps the dense parts — attention, shared experts and embeddings — in RAM, stores routed experts on disk and reads them through a cache that learns usage patterns; when a GPU is present, it holds the hottest experts and the dense layers.

The project ships 13 engines covering models including GLM-5.2/5.3, Kimi K3, DeepSeek V4 Flash (284B), MiMo-V2.6 Flash and Qwen3.6. It starts on as little as 8 GB of RAM, and the smallest model needs 22 GB of disk space. Its published measurements: GLM-5.2 runs at 1.83 tokens per second on a 128 GB Ryzen AI Max+ 395 and 9.0–9.2 tokens per second on six RTX 5090 cards. colibri also offers OpenAI- and Anthropic-compatible endpoints and an MCP server.

Key facts

1. Traction: 41,248 stars and 4,545 forks on GitHub at the time of checking, on the trending chart.
2. Resident size: GLM-5.2 (744B parameters, int4) runs with about 9.9 GB resident, activating roughly 40B parameters per token.
3. Hardware floor: 8 GB of RAM minimum, GPU optional; CUDA, Vulkan and Apple Silicon backends are supported.