prithivMLmods/Q3.6-27B-GLM-5.1-DA
prithivMLmods/Q3.6-27B-GLM-5.1-DA is a 27 billion parameter language model built on Qwen/Qwen3.6-27B, optimized for rich, detailed, and context-aware reasoning. It utilizes GLM-5.1 distilled reasoning traces and ablation-based training to reduce internal refusal behaviors while preserving strong mathematical and logical reasoning capabilities. This model excels in complex multi-step reasoning tasks and instruction-following, making it suitable for advanced research in reasoning and alignment.
Loading preview...
Q3.6-27B-GLM-5.1-DA: Reasoning-Focused Model with Reduced Refusal
Q3.6-27B-GLM-5.1-DA is a 27 billion parameter model developed by prithivMLmods, built upon the Qwen3.6-27B architecture. It is specifically optimized for enhanced reasoning, particularly in mathematical and logical contexts, by incorporating distilled reasoning traces from GLM-5.1. A key differentiator is its "Distilled-Abliterated" (DA) approach, which uses refusal direction analysis and ablation-based strategies to minimize internal refusal behaviors while maintaining high-quality reasoning and instruction-following.
Key Capabilities
- GLM-5.1 Distillation: Leverages reasoning traces from GLM-5.1 for superior mathematical and logical problem-solving.
- Reduced Refusal: Engineered to generate more direct and unrestricted content by minimizing internal refusal mechanisms.
- Strong Instruction Following: Seamlessly handles complex instructions alongside multi-step reasoning tasks.
- Qwen3.6 Backbone: Benefits from the robust performance of the Qwen3.6-27B base model.
Good For
- Mathematical Reasoning Tasks: Ideal for deep, multi-step math problems and structured thinking.
- Instruction Following: Suitable for hybrid prompts requiring both precise instruction adherence and complex reasoning.
- Red-Teaming & Alignment Research: Valuable for evaluating systems with reduced safety alignments and studying refusal mechanisms.
- Research on Abliteration: Useful for exploring the effects of ablation-based training on reasoning preservation.
Important Note: This model is intended strictly for research and learning. Due to its intentionally minimized safety refusals, it may generate sensitive or unrestricted content. Users are responsible for its ethical deployment.