llmfan46/Qwen3.6-27B-uncensored-heretic-v1

Model Overview

llmfan46/Qwen3.6-27B-uncensored-heretic-v1 is a 27 billion parameter model based on the Qwen3.6 architecture, specifically modified to reduce content refusals. This "decensored" version was created using the Heretic v1.2.0 tool with the Arbitrary-Rank Ablation (ARA) method, targeting attn.o_proj and attn.out_proj components.

Key Differentiators

• Significantly Reduced Refusals: Achieves 10 refusals out of 100, a substantial reduction compared to the original Qwen3.6-27B's 92 refusals, making it suitable for less restricted content generation.
• Preserved Model Quality: Maintains a low KL divergence of 0.0286 from the original model, indicating that its core capabilities and knowledge base are largely intact despite the decensoring process.
• Qwen3.6 Base Features: Inherits the Qwen3.6 model's enhancements in agentic coding, improved frontend workflows, repository-level reasoning, and a "thinking preservation" option for retaining reasoning context across messages.
• Multimodal Capabilities: Supports text, image, and video inputs, with a native context length of 262,144 tokens, extensible up to 1,010,000 tokens using YaRN scaling.

Performance

While the decensoring process slightly impacts MMLU accuracy (84.80% for Heretic vs 86.65% for Original), the model retains strong performance across various benchmarks, particularly in coding agent tasks (e.g., SWE-bench Pro at 53.5%) and knowledge-based tasks (MMLU-Redux at 93.5%). It also demonstrates robust vision-language capabilities, including STEM & Puzzle, General VQA, Document Understanding, and Spatial Intelligence.

Good for

• Applications requiring an LLM with fewer content restrictions.
• Agentic coding tasks, including frontend development and repository-level reasoning.
• Multimodal applications involving text, image, and video analysis or generation.
• Scenarios benefiting from extended context lengths and preserved reasoning chains.