Community
capy-flash v0.7.1 runs Qwen3.8-Flash-Next 125B MoE in 64 GB of RAM
The open-source inference engine now runs the 125B MoE on one or two 16 GB RTX 50 cards, reaching 121 tok/s on one GPU and 234 tok/s on two in the author's tests. Qwen3.8-27B at 3-bit fits on one 16 GB card at 174 tok/s, with MMLU-Pro dropping from 0.78 to about 0.72.