jaromer/JonathanColetti-Qwen3.8-27B-Uncensored

VISIONPricing:Input $1.6 / Cached $0.15 / Output $12Concurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 25, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

JonathanColetti-Qwen3.8-27B-Uncensored is a 27 billion parameter Qwen3.8-27B model developed by JonathanColetti, featuring a significantly reduced refusal rate compared to its base model. This model retains the original architecture, including vision capabilities and a 262144 token context length, but has been modified using the Heretic method to minimize refusal behavior on harmful prompts. It is primarily intended for local inference and research where reduced refusal behavior is desired, while maintaining core capabilities with a minimal performance delta.

Loading preview...

Overview of JonathanColetti-Qwen3.8-27B-Uncensored

This model is a modified version of the 27 billion parameter Qwen3.8-27B, developed by JonathanColetti, with its refusal behavior substantially reduced. It maintains the original Qwen3.8 architecture, including vision capabilities and a large 262144 token context window, operating at bf16 precision. The core difference lies in its "uncensored" nature, achieved by removing refusal directions using the Heretic method, which co-minimizes refusal count against KL divergence from the base model without fine-tuning or additional training data.

Key Capabilities & Modifications

  • Reduced Refusal Behavior: The model's refusal rate on 100 held-out harmful prompts was reduced from 98/100 in the base model to 12/100 in this version. This was achieved by modifying only attn.o_proj and mlp.down_proj modules.
  • Preserved Architecture: The model retains the Qwen3_5ForConditionalGeneration architecture, including the multi-token prediction (MTP) head and vision capabilities, which were verified and grafted back from the base checkpoint.
  • Minimal Performance Impact: Benchmarks show a mean performance delta of -0.5 across MMLU, ARC-Challenge, HellaSwag, and Winogrande compared to the base model, indicating that the refusal reduction had a minor impact on general reasoning abilities. These 0-shot scores are directly comparable between the base and uncensored versions.

Good for

  • Local Inference and Research: Ideal for users requiring a powerful 27B model with significantly reduced refusal behavior for local experimentation and research purposes.
  • Exploring Uncensored Generative AI: Suitable for use cases where the base model's safety mechanisms are overly restrictive and a more permissive response generation is desired, particularly for harmful prompts.
  • Understanding Model Behavior: Useful for studying the effects of refusal removal techniques like Heretic on large language models, especially concerning the trade-off between refusal rate and KL divergence from the base model.

Finetunes
1 models