Overview
MuXodious/Qwen2.5-7B-Instruct-1M-Thinking-Claude-Gemini-GPT5.2-DISTILL-PaperWitch-heresy is a 7.6 billion parameter instruction-tuned model based on Qwen2.5-7B-Instruct-1M. It has been fine-tuned using P-E-W's Heretic (v1.2.0) ablation engine with Magnitude-Preserving Orthogonal Ablation. This model incorporates three high-reasoning fine-tuning datasets from Claude Opus 4.5, Gemini, and GPT5.2, along with a 1 million context window, to convert the base model into a "thinking/reasoning" model.
Key Capabilities
- Enhanced Reasoning: Produces compact and "to the point" reasoning blocks, which significantly improve the quality of final output generation.
- High-Quality Output: Directly improves output generation in terms of detail, length, complexity, and overall quality.
- Temperature Robustness: Thinking/Reasoning activation is not significantly affected by temperature settings, allowing for a wide range (0.1 to 2.5 or higher).
- Self-Generating Thinking Tags: Does not require a system prompt; thinking tags/blocks will self-generate.
- Optimized for Specific Quants: Suggested quants are Q4KS (non-imatrix) or IQ3_M (imatrix) or higher for optimal reasoning activation.
Good For
- Complex Problem Solving: Excels in tasks requiring deep analytical processing and structured reasoning.
- Creative Content Generation: Recommended for creative tasks with a minimum temperature of 1.2 or higher.
- Detailed Output Generation: Ideal for scenarios where detailed, complex, and high-quality responses are crucial.
- Applications Requiring Consistent Reasoning: Suitable for use cases where consistent and compact reasoning is preferred over verbose explanations.
Note: For optimal performance, a minimum temperature of 0.7 (1.2+ for creative tasks) and a repetition penalty of 1.05 (adjustable to 1 for longer reasoning blocks) are suggested. A minimum context window of 4k, preferably 8k+, is also recommended. If thinking does not activate, users can regenerate or prepend prompts with "Think Deeply: ".