shaffhausenquant/OpenThinker-32B
OpenThinker-32B is a 32.8 billion parameter language model developed by shaffhausenquant, fine-tuned from Qwen2.5-32B-Instruct. It leverages the OpenThoughts-114k dataset, specifically designed to enhance reasoning capabilities. This model demonstrates strong performance in mathematical reasoning (MATH500: 90.6%) and general question answering (GPQA Diamond: 61.6%), making it suitable for complex analytical tasks. Its 32768-token context length supports processing extensive inputs for detailed problem-solving.
Loading preview...
OpenThinker-32B: Enhanced Reasoning Model
OpenThinker-32B is a 32.8 billion parameter language model developed by shaffhausenquant, built upon the Qwen2.5-32B-Instruct architecture. Its core differentiator is the fine-tuning process using the proprietary OpenThoughts-114k dataset, which is derived from distilling DeepSeek-R1. This specialized training aims to significantly improve the model's reasoning and problem-solving abilities.
Key Capabilities and Performance
OpenThinker-32B exhibits strong performance across several challenging benchmarks, particularly in reasoning-intensive domains:
- Mathematical Reasoning: Achieves 90.6% on MATH500, outperforming other 32B models like LIMO-32B and DeepSeek-R1-Distill-Qwen-32B.
- Complex QA: Scores 61.6% on GPQA Diamond, indicating robust general knowledge and reasoning for difficult questions.
- Overall Reasoning: Demonstrates competitive results on AIME24 I/II (66.0%) and LCBv2 (68.9%).
Training and Open-Source Commitment
The model was fine-tuned for 3 epochs with a 16k context length using LlamaFactory on AWS SageMaker with 8xH100 P5 nodes. A key aspect of OpenThinker-32B is its commitment to open-source principles, with its model weights, datasets, data generation code, evaluation code (Evalchemy), and training configurations all publicly available. This transparency allows for reproducibility and further research.
Intended Use Cases
OpenThinker-32B is well-suited for applications requiring advanced reasoning, mathematical problem-solving, and complex question answering. Its strong benchmark performance suggests utility in educational tools, research assistants, and analytical platforms where precise and logical outputs are critical.