Wickedlizerd/Noct-Q-V2-Base-NVFP4-ComfyUI

Noct Q V2 Base — native ComfyUI NVFP4

Built with Qwen. This is a community quantization of noctaluna's Noct Q V2.0 Base, not an official release by Qwen or noctaluna.

Original model: https://civitai.com/models/2958896?modelVersionId=3353641

Modified file: noct_q_v2_base_nvfp4.safetensors converts the original BF16 diffusion transformer to native ComfyUI NVFP4. All 192 attention/MLP projections are quantized; the other 73 tensors remain byte-identical BF16. No fine-tuning, merging, or changes to the text encoder or VAE were performed in this conversion.

This is the full-step Base model, not Turbo. The source author recommends 25–50 steps, CFG 1, Euler. Use the Qwen Image 2.1 encoder and 64-channel VAE. Native NVFP4 inference requires compatible ComfyUI/comfy-kitchen and Blackwell hardware. This file is intended for the native ComfyUI loader, not Diffusers.

Provenance

• Source file ID: 3241225, Noct Q V2.0 Base BF16
• Source SHA-256: 89f4158d066cc33906a199fca85634f766892dd78f49b6698dabf187ac86c4bc
• Output SHA-256: 899fd53851abe4dfe03a1ba1e4396f4e2e13bb288fbe84f2e1a067422a18d52e
• Output bytes: 4197575272
• Converter: Comfy-Org/comfy-quants, commit 3d057e9c8e3132ef1853dec96f91d3de0aaf0e6a
• Conversion: CPU, PyTorch 2.10.0, deterministic NVFP4 exporter with explicit Qwen 2.1 layer selection

Validation limits

Source checksum, output tensor structure, scale finiteness, and preservation of all BF16 tensors were checked. The converter passed 15 upstream CPU tests and four parity tests against Comfy-Kitchen reference primitives. Native Blackwell inference, output quality, and performance have not been validated.

License

The Qwen Research License applies; see LICENSE and Notice. Non-commercial use only unless separately licensed by the original rights holder. The source model author permits derivatives under the same license.