huihui-ai/Huihui-Qwen3-VL-4B-Instruct-abliterated

VISIONConcurrency Cost:1Model Size:4BQuant:BF16Ctx Length:32kPublished:Oct 16, 2025License:apache-2.0Architecture:Transformer0.1K Open Weights Cold

The huihui-ai/Huihui-Qwen3-VL-4B-Instruct-abliterated model is a 4 billion parameter vision-language instruction model, based on the Qwen3-VL architecture. This version has undergone 'abliteration' to remove refusal behaviors in its text generation capabilities, making it an uncensored variant of the original Qwen/Qwen3-VL-4B-Instruct. It retains its vision capabilities while offering unrestricted text outputs, suitable for research and experimental use where content filtering is not desired.

Loading preview...

Overview

This model, huihui-ai/Huihui-Qwen3-VL-4B-Instruct-abliterated, is a 4 billion parameter vision-language instruction model derived from the Qwen3-VL architecture. Its primary distinction is the application of 'abliteration' to the text generation component, effectively removing refusal behaviors and making it an uncensored version of the original Qwen/Qwen3-VL-4B-Instruct. The image processing capabilities remain unchanged.

Key Capabilities

  • Uncensored Text Generation: The model has been modified to eliminate responses like "I can’t describe or analyze this image," providing unrestricted text outputs.
  • Vision-Language Integration: Retains the ability to process and respond to image inputs, combining visual understanding with text generation.
  • Ollama Support: Directly usable with Ollama (version v0.12.7 or later) via ollama run huihui_ai/qwen3-vl-abliterated:4b-instruct.

Usage Warnings and Considerations

Users should be aware that the model's safety filtering has been significantly reduced. This means it may generate sensitive, controversial, or inappropriate content. It is explicitly stated as not suitable for all audiences or applications requiring high security. The developers recommend its use for research, testing, or controlled environments, advising against direct use in production or public-facing commercial applications. Users are solely responsible for ensuring their usage complies with legal and ethical standards and for monitoring model outputs.

Finetunes
2 models