Colibri is a lightweight inference engine written in pure C that allows running massive Mixture-of-Experts models (up to 2.8T parameters) on consumer-grade hardware. It eliminates external engine dependencies by implementing AI memory multitiering, treating storage, RAM, and VRAM as a unified hierarchy to stream experts from disk. The project currently supports nine major model families, including GLM-5.3, Kimi K3, and DeepSeek V4 variants, via a single C file per model.
- Zero dependency C engine reduces deployment friction and binary size for edge or local inference.
- Memory multitiering enables running 744B-2.8T parameter MoE models on hardware without massive VRAM.
- Supports diverse architectures like GLM-5.3, Kimi K3, and DeepSeek V4 via a unified CLI/web interface.
- Single C file per model simplifies integration and reduces maintenance overhead for custom deployments.
