trohrbaugh/Qwen3.6-27B-heretic
The trohrbaugh/Qwen3.6-27B-heretic is a 27 billion parameter causal language model with a vision encoder, based on the Qwen3.6 architecture developed by Qwen. This model is a decensored version of the original Qwen3.6-27B, created using Heretic v1.2.0+, and is optimized for agentic coding tasks, handling frontend workflows and repository-level reasoning with enhanced fluency. It features a native context length of 262,144 tokens, extensible up to 1,010,000 tokens, and excels in both language and vision-language benchmarks, particularly in coding agent performance and multimodal understanding.
Loading preview...
trohrbaugh/Qwen3.6-27B-heretic: Decensored Qwen3.6 for Agentic Coding
This model is a 27 billion parameter causal language model with a vision encoder, derived from the Qwen3.6 series by Qwen. It stands out as a decensored version of the original Qwen/Qwen3.6-27B, processed using Heretic v1.2.0+, which significantly reduces refusals (9/100 compared to 99/100 for the original).
Key Capabilities & Features
- Decensored Output: Offers significantly fewer refusals compared to the base model, enabling broader application contexts.
- Agentic Coding: Enhanced for complex coding tasks, including frontend workflows and repository-level reasoning, with improved fluency and precision.
- Thinking Preservation: Features an option to retain reasoning context from historical messages, streamlining iterative development and reducing overhead.
- Multimodal: Supports both text and image/video inputs, functioning as a Vision Language Model.
- Extended Context Window: Natively handles 262,144 tokens, extensible up to 1,010,000 tokens using YaRN scaling techniques.
- Strong Performance: Achieves competitive scores across various benchmarks, including SWE-bench (77.2% verified), MMLU-Redux (93.5%), and MMMU (82.9%).
What Makes This Different?
Unlike the base Qwen3.6-27B, this "heretic" variant is specifically modified to be decensored, making it suitable for use cases where the original model's refusal rates might be prohibitive. It maintains the robust agentic coding and multimodal capabilities of the Qwen3.6 family while offering greater flexibility in content generation.
Should You Use This?
- Yes, if: Your application requires a powerful 27B multimodal model with strong coding and reasoning abilities, and you need a model with significantly reduced content refusal rates. It's particularly well-suited for agentic applications, complex coding tasks, and scenarios benefiting from extended context and thinking preservation.
- Consider alternatives if: You require strict content moderation or prefer a model with default safety alignments, as the decensored nature of this variant means it will generate content that the original Qwen3.6-27B might refuse.