crit-hub

crit-hubrtx5090Qwen/Qwen3.8-27Bunsloth/Qwen3.8-27B-NVFP4

EngineContributorBenchmarkGreedy decodeSampled decodePrefillTTFTPerplexityKL divergenceContext
mtp7-device-confidencecrit-labscrit-eval-v6252.4 tok/s–7,425 tok/s150 ms+0.1%0.0028262,144 tok