Model Overview
DavidAU/Qwen2.5-7B-Instruct-1M-Thinking-Claude-Gemini-GPT5.2-DISTILL is a 7.6 billion parameter model based on Qwen2.5-7B-Instruct-1M, enhanced through Unsloth fine-tuning. Its core innovation lies in converting the base model into a "thinking/reasoning" model by integrating three high-reasoning fine-tuning datasets from Claude Opus 4.5, Gemini, and GPT5.2. This process enables the generation of compact and "to the point" reasoning blocks, which significantly improve the quality, detail, length, and complexity of the final output.
Key Capabilities
- Enhanced Reasoning: Utilizes a unique, compact reasoning/thinking "engine" derived from advanced LLM datasets.
- Improved Output Quality: Reasoning directly enhances the quality of generated text, including detail, length, and complexity.
- Flexible Temperature Settings: Reasoning activation is not affected by temperature, allowing for a wide range (0.1 to 2.5+) for creative outputs.
- Dynamic Thinking Activation: Reasoning blocks adapt in depth and complexity based on prompt complexity; can be explicitly activated with "Think Deeply:" prefix.
Recommended Usage
- Settings: Suggested parameters include
temp .7+ (1.2+ for creative), rep pen 1.05 (1.1-1.15 with smoothing), topp .95, minp .05, topk 40. - Context Window: Minimum 4k, but 8k+ is suggested for optimal performance.
- Quantization: Q4KS (non-imatrix) or IQ3_M (imatrix) or higher are recommended to avoid reasoning issues.
- Smoothing Factor: For chat/roleplay, setting a "Smoothing_factor" to 1.5 in interfaces like KoboldCpp, oobabooga, or Silly Tavern is advised for smoother operation.