Description
Frontier models from 744B to 2.8T parameters normally need a multi-GPU server. colibri, in pure C with zero dependencies, treats storage, RAM and VRAM as one inference hierarchy, streaming experts from disk so consumer and mixed hardware can run them.
Nine model families including GLM-5.2 and 5.3 run today on a tiny engine.
Huge models:744B to 2.8T MoE.
Memory tiers:Disk, RAM and VRAM unified.
Pure C:Small and portable.
Model families:Nine supported.
Nine model families including GLM-5.2 and 5.3 run today on a tiny engine.
Features
Huge models:744B to 2.8T MoE.
Memory tiers:Disk, RAM and VRAM unified.
Pure C:Small and portable.
Model families:Nine supported.
