Anyone curious enough should be able to open one up — run it, watch every expert fire, and make it better.
Read the engine →A token through GLM-5.2 activates only ~40B parameters, and just ~11 GB of those change from token to token — the routed experts. colibri keeps the dense weights resident and treats the 19,456 experts as a storage hierarchy: measured routing heat decides which experts earn VRAM, which earn pinned RAM, and which stream from NVMe. The router runs a layer ahead, so prefetch hides the latency.
Everything is validated token-exact against the reference transformers implementation — speed never buys drift.
The flagship target: an int4 container, token-exact versus reference, with the full expert atlas published.
Run it →The small research workhorse — quantization A/Bs and quality ablations run here first.
Run it →A 975B reasoning MoE, running today. An int4 dense container brings the resident set to 15.3 GB, so it fits a 25 GB box.
Run it →A widely-studied frontier MoE family — the tiered engine's next large-scale target.
On the roadmapThe largest model here, running today — its QAT-trained MXFP4 experts are streamed straight from the original checkpoint, with no conversion.
Run it →The most widely-deployed open family — broad hardware coverage meets broad adoption.
On the roadmapEvery point behind this text is a real measured expert — 13,260 characterised experts, specialists clustering by topic, generalists drifting to the centre. The first expert atlas of a 700B-class model, mapped on ordinary machines.
Measured decode speed on real community machines. The hardware only changes where the experts live, not what the model answers.
…and 15+ more community machines, from an M4 Pro Mac Mini to a 430 GB EPYC server. Full tables, methodology and quality ablations: docs/benchmarks.md