Qwen-3.5-Abliterated-Comfyui-nvfp4 unlocks local AI power

    
        By vramkickedin    
     | 
    
            April 7, 2026 at 2:09 am        
    
     | 
    
        2 min read

Qwen-3.5-Abliterated-Comfyui-nvfp4 is a collection of quantized language models designed to function as AI assistants directly within ComfyUI. Developer Winnougan created these models to enable multimodal tasks like image analysis and prompt generation without relying on external tools or cloud services.

The release offers "abliterated" variants that remove built-in refusal behaviors from the original Qwen 3.5 models while maintaining their core capabilities. Users can select from multiple quantization formats, including NVFP4 and MXFP8, allowing the models to run efficiently on consumer hardware.

Model Size: from 3.54GB & VRAM GPU: requirements vary

Core capabilities and abliterated integration

Native ComfyUI compatibility using standard CLIP Loader nodes.
Multimodal image understanding and detailed captioning.
Multiple quantization options ranging from 3.54 GB to 11.9 GB files.
Abliteration technique removes refusal behavior without retraining.
Prompt generation for video models like LTX and WAN.
Both 4B and 9B parameter variants available.
Local processing maintains user privacy.

Creative professionals working with image and video generation pipelines can use these models to generate detailed prompts, analyze existing images, or automate workflow tasks. The abliteration process allows for unrestricted outputs, making the models suitable for creative writing and experimental applications where standard safety filters might interfere with legitimate use cases.

Technical notes from the developer

Winnougan released these models as community-created variants, noting that they are not official releases from the original Qwen team. The NVFP4 quantization format provides the smallest file sizes and fastest inference but requires newer NVIDIA GPU architectures for optimal performance.

Users should note that abliterated models remove alignment layers designed to prevent harmful outputs.