ComfyUI isn't just for images: it runs local language models for prompt enhancement, image descriptions, stories and basic chat. In a blog post on 11 November 2026, the Comfy team said a rewrite of that text-generation path, focused on the Qwen3.x family, lets the same checkpoints on the same GPU generate tokens up to twice as fast, with no change to the model or the quality of what it writes. This guide explains how to get the speed-up, based on Comfy's blog and documentation; Forge PC hasn't benchmarked it on our own hardware.
What changed
Comfy's blog says everything it describes is available in ComfyUI v0.45.0. The main points:
Multi-token prediction (MTP). Qwen3.x checkpoints ship with an MTP head, a smaller extra layer that guesses upcoming tokens. ComfyUI uses those guesses for speculative decoding: it drafts a few tokens, then checks them all in one pass. Every correct guess saves a full read of the model.
Adaptive tree MTP. Instead of betting on a single chain of guesses, ComfyUI keeps a few runner-up tokens at each step and re-plans the shape of its guesses as it goes, based on what's been hitting.
Faster kernels. Comfy rewrote the matrix and attention kernels used for single-sequence text generation to make better use of VRAM bandwidth.
Why does that help? Comfy explains that generating text is limited by how fast your GPU can read the model from VRAM, not raw compute. For a 27B model quantised to four bits, Comfy says each step reads roughly 14GB of weights to produce one token, which is why checking several guessed tokens in the same pass is, in Comfy's words, "almost free".
Comfy's own numbers
These are Comfy's figures, not ours:
Model and task (Comfy's test) | GPU | Before | After |
|---|
Qwen3.5 9B | RTX 5090 | 106 tokens/s | 219 tokens/s |
Qwen3.5 9B | RTX 5080 | 91 tokens/s | 129 tokens/s |
Qwen3.8 27B describing a 1536×1536 image | RTX 5090 | 55 tokens/s | 122 tokens/s |
Comfy says the improvements run on NVIDIA and AMD GPUs, on Windows and Linux. The size of any speed-up will depend on your GPU, the model and the kind of text being written. Comfy notes that predictable output such as code and lists suits long chains of guesses, while creative writing suits wider trees.
Step 1: Update ComfyUI to v0.45.0 or later
How you update depends on how you installed it, according to ComfyUI's update guide:
ComfyUI Portable (Windows): in the update folder, run update_comfyui.bat for the latest development version, or update_comfyui_stable.bat for the latest stable release. Comfy warns against using update_comfyui_and_python_dependencies.bat for routine updates, because it reinstalls all Python dependencies and can break custom nodes.
Comfy Desktop: open the Manage panel and use the Update tab, choosing the Stable (default) or Latest on GitHub channel. The Desktop app itself updates separately and automatically.
Manual (Git) installs: activate ComfyUI's virtual environment, run git pull in the install folder, then pip install -r requirements.txt, and restart with python main.py.
If you're new to ComfyUI, start with our guide to installing ComfyUI on Windows.
Step 2: Load a Qwen3.x checkpoint and set MTP to auto
Comfy's instructions are short: load a Qwen3.x checkpoint, leave the Text Generate node's mtp option on auto, and run. Comfy says adaptive tree MTP is used when mtp is set to auto. Comfy says the improvements don't require additional steps beyond an up-to-date ComfyUI.
Step 3 (optional): Try Comfy-Kitchen attention
Comfy says ComfyUI v0.45 adds INT8 attention with KV cache support for these models, which reduces the KV cache's VRAM footprint and speeds up pre-fill for long prompts. You enable it by launching ComfyUI with:
--use-ck-attention
Comfy also says longer contexts that used to run out of memory now fit. If you're short of VRAM, our guide to running ComfyUI on low VRAM has more options.
Our verdict
If you use ComfyUI's Text Generate node with Qwen3.x models, update to v0.45.0 or later and leave mtp on auto: by Comfy's own figures that can roughly double token rates on the same GPU. Your gains will vary with hardware and model, so measure before and after rather than relying on headline numbers.