JonathanColetti/Qwen3.8-27B-Uncensored
JonathanColetti/Qwen3.8-27B-Uncensored is a 27 billion parameter Qwen3_5ForConditionalGeneration model, derived from Qwen/Qwen3.8-27B, with its refusal behavior substantially reduced using the Heretic method. This bf16 model retains the original architecture, including vision capabilities and a 262144-token context length, while aiming to minimize refusals on harmful prompts. It is primarily intended for local inference and research where reduced refusal behavior is desired, with a measured 88% reduction in refusals on a specific harmful prompt dataset.
Loading preview...
Model Overview
JonathanColetti/Qwen3.8-27B-Uncensored is a 27 billion parameter model based on the Qwen/Qwen3.8-27B architecture, specifically Qwen3_5ForConditionalGeneration. Its primary distinction is the substantial reduction of refusal behavior on harmful prompts, achieved through the Heretic method. This process co-minimizes refusal count against KL divergence from the base model, without additional fine-tuning or training data.
Key Characteristics & Performance
- Reduced Refusals: Achieves a reduction from 98/100 refusals to 12/100 on a set of 100 held-out harmful prompts, representing an 88% decrease in refusal rate for these specific prompts.
- Architecture: Retains the original Qwen3.8-27B architecture, including 64 layers, a 248320-token vocabulary, and vision capabilities.
- Context Length: Supports a large context of 262144 tokens.
- Precision: Provided in bf16 precision, requiring approximately 55 GB of VRAM for local inference.
- 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, with individual deltas largely within reported standard error.
- MTP Head Present: The multi-token prediction (MTP) head is verified to be present and functional.
Intended Use Cases
- Local Inference and Research: Ideal for developers and researchers conducting local experiments where a less restrictive model is preferred.
- Exploration of Uncensored Behavior: Useful for studying model responses to prompts that the base model would typically refuse.
Limitations
- Refusals are reduced, not eliminated, and are not redirected. A meaningful fraction of requests may still be refused.
- The model represents the highest-divergence point on the Pareto front, meaning behavior near the old refusal boundary may be less stable than the base model.
- Capability benchmarks did not cover generative, math, code, or multilingual evaluations, nor the vision tower or MTP speculative decoding.