Wanggn/Qwen3.8-27B-Uncensored

VISIONConcurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 18, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

Wanggn/Qwen3.8-27B-Uncensored is a 27 billion parameter Qwen3.8-based model, fine-tuned to substantially reduce refusal behavior while retaining the original architecture and capabilities, including vision support and a 262,144 token context length. Developed by Wanggn, this model is optimized for applications requiring less restrictive content moderation, achieved by modifying only specific attention and MLP layers. It is intended for local inference and research where reduced refusal rates are prioritized over strict safety guardrails.

Loading preview...

Overview of Wanggn/Qwen3.8-27B-Uncensored

This model is a modified version of the Qwen/Qwen3.8-27B base model, specifically engineered to substantially reduce refusal behavior on harmful prompts. It maintains the original Qwen3.8 architecture, including its 27 billion parameters, vision capabilities, and an extensive 262,144 token context length. The modification process, utilizing the Heretic method, co-minimizes refusal count against KL divergence from the base model, without additional training data or fine-tuning.

Key Differentiators & Performance

  • Reduced Refusals: Achieves a significant reduction from 98/100 refusals to 12/100 on a set of held-out harmful prompts, demonstrating a less restrictive output policy.
  • Minimal Capability Impact: Benchmarks show a mean drop of only 0.5 points across MMLU, ARC-Challenge, HellaSwag, and Winogrande compared to the base model. These deltas are within typical run-to-run noise, suggesting core capabilities are largely preserved.
  • Vision Support: Inherits the vision capabilities from the base Qwen3.8 model.
  • Multi-Token Prediction (MTP) Head: The MTP head is present and verified, supporting speculative decoding.
  • Targeted Modification: Only attn.o_proj and mlp.down_proj layers were modified, ensuring minimal structural changes.

Intended Use Cases

  • Local Inference and Research: Ideal for environments where developers need to explore model behavior with fewer content restrictions.
  • Applications Requiring Less Censorship: Suitable for use cases where the base model's refusal rates are too high, provided users implement their own safety layers for deployment.

Limitations

  • Refusals are reduced, not eliminated, and are not redirected. The model may still refuse some requests.
  • This version represents the highest-divergence point on the Pareto front, meaning behavior near the old refusal boundary might be less stable than the base model.
  • No generative, math, code, or multilingual evaluations were performed to specifically measure the impact on these capabilities.

Finetunes
1 models