# ComfyUI v0.45.0 doubles Qwen3.x local text generation speed

> Original title: Doubling LLM token rates in ComfyUI

- Source: Comfy Blog
- Published: 2026-11-11T16:56:59.000Z
- Bad Influence: https://badinfluence.news/items/tqyjhd8f7auv83esbbdmczh2n
- Original: https://blog.comfy.org/p/doubling-llm-token-rates-in-comfyui

## Summary

ComfyUI says its v0.45.0 update roughly doubles Qwen3.x token generation speed on the same hardware and checkpoints, without changing model quality. On an RTX 5090, Qwen3.5 9B rose from 106 to 219 tokens per second, and Qwen3.8 27B describing a 1536×1536 image rose from 55 to 122. The gains come from a rewritten linear kernel, adaptive tree multi-token prediction, and a new attention path, and the post says they run on NVIDIA and AMD without extra setup.

## Why it matters

The post explains how ComfyUI's kernel, MTP, and attention changes raise Qwen3.x token rates, useful for judging local inference performance tradeoffs.