Colibrì Runs 744B- to 2.8T-Parameter MoE Models on 25 GB Machines by Streaming Experts From Disk

Colibrì now runs MoE models as large as Kimi K3’s 2.8 trillion parameters on machines with just 25 GB of memory by streaming routed experts from disk, though cold GLM-5.2 inference falls to 0.05–0.1 tokens per second.

• Colibrì runs eight frontier MoE model families, from 744B-parameter GLM-5.2 to 2.8T-parameter Kimi K3, using a pure-C engine that streams experts from disk.
• Its GLM-5.2 int4 container stores 19,456 routed experts across about 372 GB while keeping the 17B-parameter dense portion resident in roughly 9.9 GB of RAM.
• On six RTX 5090 GPUs with all experts resident, Colibrì reports 5.8–6.8 tokens per second; a 25 GB machine runs the same model at 0.05–0.1 tokens per second cold.