llmfan46/Qwen3.6-27B-uncensored-heretic-v2

Overview

This model, llmfan46/Qwen3.6-27B-uncensored-heretic-v2, is a 27 billion parameter decensored version of the Qwen/Qwen3.6-27B causal language model. It was created by llmfan46 using the Heretic v1.2.0 tool, employing a variant of the Magnitude-Preserving Orthogonal Ablation (MPOA) method to reduce content refusals. The model achieves a 94% reduction in refusals (6/100 compared to 92/100 for the original) with a minimal KL divergence of 0.0021, indicating strong preservation of the original model's quality.

Key Capabilities & Features

• Decensored Output: Significantly fewer content refusals (94% reduction) compared to the base Qwen3.6-27B model, making it suitable for use cases requiring less restrictive content generation.
• Agentic Coding: Enhanced capabilities in handling frontend workflows and repository-level reasoning, as highlighted by strong performance on benchmarks like SWE-bench and Terminal-Bench 2.0.
• Thinking Preservation: Features an option to retain reasoning context from historical messages, which streamlines iterative development and can reduce token overhead.
• Multimodal: Supports both text and image inputs, and is a causal language model with a vision encoder.
• Extended Context Length: Natively supports a context length of 262,144 tokens, extensible up to 1,010,000 tokens using YaRN scaling techniques.

Performance Highlights

While primarily focused on decensoring, the model maintains strong performance across various benchmarks:

• MMLU Accuracy: Achieves 85.61% on MMLU (Massive Multitask Language Understanding), a slight decrease from the original's 86.65%, demonstrating quality preservation.
• Coding Agent Benchmarks: Shows competitive results in coding agent tasks, including SWE-bench Verified (77.2), SWE-bench Pro (53.5), and Terminal-Bench 2.0 (59.3).
• Vision Language Benchmarks: Performs well in multimodal tasks such as MMMU (82.9) and MathVista mini (87.4).

Use Cases

This model is particularly well-suited for:

• Applications requiring less restrictive content generation: Where the base model's refusal rates are too high.
• Advanced Coding Assistance: For developers needing robust support for complex coding tasks, including frontend development and repository-level reasoning.
• Agentic Workflows: Leveraging its agentic coding and thinking preservation features for more efficient and consistent automated tasks.
• Multimodal Applications: Integrating text and image inputs for diverse use cases.