--- license: other license_name: qwen-research license_link: LICENSE base_model: - Qwen/Qwen-Image-2.1 tags: - comfyui - nvfp4 - image-to-image - quantized --- # 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.