EzekielBlaze/Nia-Qwen3-VL-32B-Thinking-eos-fix

VISIONConcurrency Cost:2Model Size:33.4BQuant:FP8Ctx Length:32kTool Calling:SupportedPublished:Jun 12, 2026License:apache-2.0Architecture:Transformer Open Weights Cold

EzekielBlaze/Nia-Qwen3-VL-32B-Thinking-eos-fix is a 33.4 billion parameter multimodal large language model, based on the Qwen3-VL-32B-Thinking architecture. This model has been uncensored using abliteration techniques, specifically targeting the text generation component to remove refusal behaviors. It excels at multimodal tasks, particularly image description and analysis, and is noted for achieving a perfect score on the UGI Leaderboard's W10 category for non-refusal to unsafe instructions.

Loading preview...

Model Overview

This model, EzekielBlaze/Nia-Qwen3-VL-32B-Thinking-eos-fix, is an uncensored variant of the Qwen3-VL-32B-Thinking 33.4 billion parameter multimodal large language model. It has been processed using abliteration techniques to remove refusal behaviors in its text generation, specifically addressing responses like "I can’t describe or analyze this image." The image processing capabilities remain unchanged from the base model.

Key Capabilities & Differentiators

  • Uncensored Text Generation: The primary differentiator is the removal of safety filtering in the text component, allowing it to respond to prompts that standard models might refuse.
  • Multimodal (Vision-Language): Capable of processing and generating text based on image inputs.
  • UGI Leaderboard Performance: Achieves a perfect score in the W10 category of the UGI Leaderboard, indicating complete non-refusal to unsafe instructions.
  • Ollama Support: Directly available for use with Ollama (version v0.12.7 or newer) via ollama run huihui_ai/qwen3-vl-abliterated:32b.

Usage Warnings & Considerations

Due to the significant reduction in safety filtering, users should be aware of:

  • Risk of Sensitive Outputs: The model may generate controversial or inappropriate content.
  • Limited Suitability: Not recommended for public-facing applications, underage users, or scenarios requiring high security.
  • User Responsibility: Users are solely responsible for the legal and ethical implications of generated content.
  • Research Use: Best suited for research, testing, or controlled environments rather than production.
  • No Safety Guarantees: Unlike standard models, this version lacks rigorous safety optimization, and the developers bear no responsibility for its use.

Finetunes
1 models