Model Overview

huihui-ai/Qwen3-4B-abliterated is an uncensored variant of the Qwen3-4B language model, developed by huihui-ai. This model utilizes an "abliteration" technique, a method for removing refusal behaviors from LLMs, as detailed in the remove-refusals-with-transformers project. The current version represents a proof-of-concept implementation with a new, faster abliteration method that reportedly yields improved results.

Key Capabilities & Differentiators

• Uncensored Output: The primary feature is its ability to generate responses without the refusal behaviors typically present in moderated LLMs.
• High Pass Rate on Harmful Instructions: The model demonstrates a 100.00% pass rate (320/320) on a specific set of harmful instruction tests, significantly outperforming the base Qwen3-4B model which achieved an 81.56% pass rate. This indicates its effectiveness in bypassing content restrictions.
• Efficient Abliteration: The model was created using a new and faster abliteration method, suggesting potential for more efficient development of uncensored models.

Usage & Availability

This model can be loaded and used with the Hugging Face transformers library. An Ollama version is also available for direct use via ollama run huihui_ai/qwen3-abliterated:4b.

Note: A newer version, huihui-ai/Huihui-Qwen3-4B-abliterated-v2, is available and recommended for use.