Community
Kapitsa 0.2.26 benchmarks Qwen and Nemotron models on a 128 GB M5 Max MacBook Pro
On a 128 GB M5 Max, Kapitsa 0.2.26 decodes Qwen3.6 35B-A3B at 346 tok/s and Qwen Flash-Next 180B at 108 tok/s, based on median of three runs. The engine also runs Flash-Next 180B on a 48 GB MacBook Pro using SSD offloading at 61 tok/s.