crit-hub

crit-hubrtx5090Qwen/Qwen3.8-Flash-Nextnvidia/Qwen3.8-Flash-Next-NVFP4

EngineContributorBenchmarkGreedy decodeSampled decodePrefillTTFTPerplexityKL divergenceContext
mtp3-host-expertsexternalgiveenself-reported · LocalMaxxing140.9 tok/s–––3.518 (absolute)–262,144 tok