llmfan46/Qwen3.6-27B-uncensored-heretic-v2-Native-MTP-Preserved
The llmfan46/Qwen3.6-27B-uncensored-heretic-v2-Native-MTP-Preserved is a 27 billion parameter causal language model based on the Qwen3.6 architecture, developed by Qwen and decensored by llmfan46. This model significantly reduces refusals by 94% (6/100 vs 92/100) while maintaining original model quality with a KL divergence of 0.0021. It is optimized for agentic coding, including frontend workflows and repository-level reasoning, and supports a native context length of 262,144 tokens, extensible up to 1,010,000 tokens.
Loading preview...
Overview
This model, llmfan46/Qwen3.6-27B-uncensored-heretic-v2-Native-MTP-Preserved, is a 27 billion parameter decensored version of the Qwen3.6-27B model. It was created using the Heretic v1.3.0 tool with a variant of the Magnitude-Preserving Orthogonal Ablation (MPOA) method, ensuring that 15 Multi-Token Predictions (MTPs) are preserved. The primary goal of this variant is to drastically reduce content refusals while maintaining the original model's quality.
Key Differentiators
- Significantly Reduced Refusals: Achieves 94% fewer refusals (6/100) compared to the original model (92/100), making it highly suitable for less restricted content generation.
- Quality Preservation: Maintains a low KL divergence of 0.0021, indicating minimal deviation from the original model's performance and capabilities.
- Enhanced Agentic Coding: The base Qwen3.6 model excels in agentic coding tasks, including frontend workflows and repository-level reasoning.
- Extended Context Length: Supports a native context length of 262,144 tokens, extensible up to 1,010,000 tokens using YaRN scaling techniques.
- Thinking Preservation: Features an option to retain reasoning context from historical messages, improving decision consistency and inference efficiency.
Performance
While significantly reducing refusals, the model maintains strong performance across various benchmarks. For instance, its MMLU accuracy is 85.67%, closely matching the original's 86.65%. It also shows competitive scores in coding agent benchmarks like SWE-bench Verified (77.2) and Terminal-Bench 2.0 (59.3), and strong vision-language capabilities across MMMU and RealWorldQA.
Use Cases
This model is particularly well-suited for applications requiring high-performance language generation with minimal content restrictions, especially in agentic coding, complex problem-solving, and multimodal tasks involving images and video.