Qwen3.8-27B-NVFP4-ComfyUI

ComfyUI node update (Sep 17, 2026):

Fix nvfp4 vision-tower crash with image/video input on ComfyUI 0.35.1+:
the vision tower now stays on the full-precision path instead of being switched
to W4A4 nvfp4 matmul (which can't take the float32 vision activations).

Qwen3.8-27B ComfyUI NVFP4

A ComfyUI-native quantized conversion of unsloth/Qwen3.8-27B-NVFP4

Unlike the core node, the custom node doesn’t run for max_length steps and finishes within seconds

<a href="https://huggingface.co/rzgar/Qwen3.8-27B-NVFP4-ComfyUI/resolve/main/workflow/screenshot.png" target="_blank"><img src="https://huggingface.co/rzgar/Qwen3.8-27B-NVFP4-ComfyUI/resolve/main/workflow/screenshot.png" width="1280" height="720" /></a>

File

Usage

Place the file in ComfyUI/models/text_encoders/ and load it with the LLM/text-encoder loader. Use with the Generate Text node.

Recommended: the "Generate Text (Qwen3.8)" custom node (install ComfyUI-Qwen3.8-Text)

The official Generate Text node for quantized checkpoints, it projects the output through

the embedding table instead of the real lm_head (untied), which produces incoherent text that never stops and runs to max_length.

The custom node formats the prompt with the model's own chat template, uses the correct lm_head weights, and supports image and video inputs typical captions/answers finish in seconds.

Comparison