Local Qwen3.x models generate up to twice as fast in ComfyUI v0.45.0
AI-generated summary Local Qwen3.x models generate up to twice as fast in ComfyUI v0.45.0 through a rewritten generation hot path that includes an optimized GEMV kernel and INT8 attention. The team states that enabling mtp=auto in the Text Generate node increases speed up to two times with no change in quality.