Qwen3-4B-Instruct-Uncensored: Abliterated for Unrestricted Responses
This model, developed by n0ctyx, is an uncensored version of the Qwen3-4B-Instruct-2507 base model. It leverages a technique called abliteration to surgically remove the internal "refusal direction" from the model's weights, eliminating safety refusals without retraining. This process modifies the weights directly through orthogonalization, preserving the original model's intelligence and capabilities.
Key Characteristics & Performance
- Base Model: Qwen3-4B-Instruct-2507 (4.0B parameters, 36 layers, 262,144 token context length).
- Abliteration: Achieves a KL Divergence of 0.0785, indicating minimal loss of original intelligence.
- Reduced Refusals: Responds to approximately 81% of previously refused prompts, with only 19/100 refusals remaining, typically on extreme edge cases.
- Architecture: Dense transformer with GQA (32 Q-heads, 8 KV-heads).
Ideal Use Cases
- Creative Writing: Generating fiction, roleplay, and character dialogue without content restrictions.
- Research & Testing: Useful for red-teaming, safety analysis, and adversarial testing of language models.
- Dataset Generation: Creating synthetic training data for fine-tuning other models.
- Unfiltered Assistance: Providing direct answers without hedging or artificial gatekeeping.
Limitations
While significantly uncensored, the model may still exhibit a 19% refusal rate on the most extreme prompts. As a 4B parameter model, it may not be optimal for highly complex reasoning tasks compared to larger variants. Users should be aware that it may occasionally produce inaccurate or hallucinated content, similar to its base model, and should be used responsibly.