A 430 GB Model Runs at 0.027 Tokens a Second, and the CPU Is Not the Problem
Colibri v2.1.0 ran eight text models and one image model on a GPU-less 12 vCPU Xeon with 47 GiB of RAM. Throughput falls from 0.646 tokens a second on a 23 GB Qwen3 model to 0.027 on a 429 GB GLM, a 24 times gap for a twenty times increase in size. The engine attributes 65 to 96 percent of request time to expert disk reads.

