Colibri is a lightweight inference engine written in pure C that supports frontier Mixture-of-Experts models ranging from 744B to 2.8T parameters. It operates without external dependencies by treating storage, RAM, and VRAM as a unified hierarchy to stream experts from disk. The tool supports eight model families, including GLM-5.3, Kimi K3, and DeepSeek V4, providing consistent CLI and web frontends across all.
- Pure C implementation ensures zero external dependencies and minimal overhead.
- Supports massive MoE models (up to 2.8T) via disk-to-VRAM expert streaming.
- Unified hierarchy manages storage, RAM, and VRAM for efficient multitiering.
- Covers eight major model families with consistent coli chat/serve/web UIs.
- Enables running frontier models on consumer or heterogeneous hardware.
