richardyoung/Qwen3.6-27B-heretic

Model Overview

This model, richardyoung/Qwen3.6-27B-heretic, is a 27 billion parameter causal language model based on the Qwen3.6 architecture, specifically modified using Heretic v1.4.0 to be a decensored version of the original Qwen/Qwen3.6-27B. It integrates a vision encoder and supports a native context length of 262,144 tokens, which can be extended up to 1,010,000 tokens using YaRN scaling techniques.

Key Differentiators

• Decensored Version: Modified using Heretic v1.4.0, resulting in a significantly lower refusal rate (38/100) compared to the original Qwen3.6-27B (91/100).
• Enhanced Agentic Coding: Demonstrates improved performance in frontend workflows and repository-level reasoning, with notable scores on benchmarks like SWE-bench Verified (77.2) and Terminal-Bench 2.0 (59.3).
• Thinking Preservation: Features an option to retain reasoning context from historical messages, which streamlines iterative development and can reduce token consumption.
• Multimodal Capabilities: Supports both image and video inputs, functioning as a Vision Language Model (VLM) with strong performance across various VQA and document understanding benchmarks.
• Extended Context Window: Natively handles 262,144 tokens and is extensible to over 1 million tokens, beneficial for complex, long-horizon tasks.

Ideal Use Cases

• Coding Agents: Excellent for applications requiring advanced code generation, debugging, and understanding large codebases, especially in agentic scenarios.
• Multimodal AI: Suitable for tasks involving image and video analysis, visual question answering, and document understanding.
• Complex Reasoning: Benefits from its thinking preservation feature and large context window for tasks demanding sustained, iterative reasoning.
• Applications Requiring Reduced Refusals: Its decensored nature makes it more permissive for a broader range of queries compared to its base model.