Sources
Colibrì runs 744B to 2.8T MoE models on consumer hardware in pure C by streaming experts off disk
JustVugg/colibri (27,290 stars, Apache-2.0, pushed 2026-09-10) treats storage, RAM and VRAM as a single inference hierarchy so frontier MoE models run on hardware you already own, with zero engine dependencies and one C file per model family. Eight families work today: GLM-5.2/5.3 at 744B, GLM-5.3-Flash at 321B with vision, Inkling at 975B, Kimi K3 at 2.8T, DeepSeek V4 Flash at 284B, Qwen3.8-Flash-Next, Qwen3.6 and OLMoE, all behind the same coli chat / coli serve / coli web front end. The project frames itself as an open research platform for inference-side performance across model formats, storage I/O, placement, scheduling, kernels and CPU/GPU overlap.
↳ Follow the thread