Grimxlock/Qwen3.8-27B-Abliterated
Grimxlock/Qwen3.8-27B-Abliterated is a 27 billion parameter multimodal language model, derived from Qwen3.8-27B, that has been specifically modified to eliminate refusal behaviors. Through targeted activation steering, this model achieves zero refusals on harmful prompts while largely preserving its original capabilities. It is designed for applications requiring a highly compliant model that avoids content restrictions, making it suitable for diverse generative tasks.
Loading preview...
Grimxlock/Qwen3.8-27B-Abliterated: Refusal-Free Multimodal LLM
This model is an "abliterated" variant of the Qwen3.8-27B multimodal language model, featuring 27 billion parameters. Its primary distinction lies in the complete removal of refusal behaviors, achieved through a technique called targeted activation steering/ablation applied across the model's internal architecture.
Key Capabilities & Differentiators
- Zero Refusals: Evaluated against 450 harmful prompts, the model demonstrated 0 refusals and 0 degenerate outputs, a significant reduction from the base Qwen3.8-27B's 283 refusals.
- Preserved Performance: Despite the ablation, the model maintains nearly all of the base model's capabilities. It passed 14 out of 15 capability benchmark tasks, with the only noted degradation being a minor error in a specific numeric/math problem.
- Multimodal: Inherits the multimodal capabilities of the base Qwen3.8-27B, allowing it to process and generate content involving various data types.
- Methodology: The abliteration process involved computing a refusal/steering direction from contrastive activations (refusal vs. compliant prompts) and then removing or steering that direction from the model weights. This is similar to the methodology used in the openbmb-MiniCPM5-1B-F16-Annihilated project.
Good For
- Use cases where strict compliance and the avoidance of refusal behaviors are paramount.
- Applications requiring a robust multimodal LLM that can generate diverse content without encountering content restrictions.
- Developers seeking a powerful 27B parameter model with high instruction following and reasoning capabilities, free from built-in refusal mechanisms.