Community NVFP4 quantization of Noctaluna's Noct Q V4.0 Base: https://civarchive.com/models/2958896?modelVersionId=3372040 . Not an official Qwen or Noctaluna release. All credit for the model goes to Noctaluna.
Uncensored like the original: nudes and explicit scenes render without a LoRA. The example images were made with this NVFP4 file and the standard qwen3vl_8b text encoder (NVFP4), no negative prompt: 80 steps, euler_ancestral_cfg_pp, beta, cfg 0.7, 1152x1728 or 1408x1408 (about 2 megapixels), captions of 300 to 360 words. The author's own setting (25 steps, euler, simple, cfg 3) also works and takes about a third of the time.
What it is: the author's BF16 diffusion transformer converted to native ComfyUI NVFP4. All 192 attention and image-MLP projections are quantized, the other tensors stay BF16. No fine-tuning was done. The file is 3.9 GB instead of 13.3 GB. The text encoder and VAE are not included.
How to use: put the file in models/diffusion_models and load it with the standard UNET loader, weight dtype default. Use the Qwen-Image 2.1 text encoder and qwen_image_2.1_vae_bf16. Native NVFP4 needs an RTX 50-series (Blackwell) card.
Check: loaded in ComfyUI on an RTX 5080, the log reported native NVFP4 ops. A 1024x1024, 20-step, Euler/simple, CFG 1 run finished in 5.76 seconds.
Source: Civitai file 3260162 (BF16), SHA-256 0e4b9285fe54571660d0b5d2559ade4bec9a350ce633ad2d31192f88f9d5e0d3. This file: SHA-256 e78632166ecbe84efa4e980b369822c9fcb3c6bf2bebf4f13acef90dbad88915. Also on Hugging Face: https://huggingface.co/sjoe1244/Noct-Q-V4-Base-NVFP4
License: Qwen Research License, non-commercial. Same permissions as the original model.
Description
NVFP4 conversion of the V4.0 Base BF16 file (Civitai file 3260162). 192 attention and image-MLP matrices are quantized, the rest stays BF16. Author's settings for V4 Base: 25 steps, euler, simple, cfg 3 (cfg 1 runs about twice as fast). Text encoder: the standard qwen3vl_8b, the same one the author's workflow loads. VAE: qwen_image_2.1_vae_bf16.