Hugging Face OpenThinker2-32B: Enhanced Reasoning Model
OpenThinker2-32B is a 32.8 billion parameter language model developed by open-thoughts, building upon the Qwen2.5-32B-Instruct architecture. It represents an advancement over the previous OpenThinker-32B, primarily through fine-tuning on the expanded OpenThoughts2-1M dataset.
Key Capabilities & Performance
This model is specifically designed and trained to excel in complex reasoning tasks, particularly in mathematics and code. Its training data includes a significant augmentation of math and code reasoning examples, generated through diverse methodologies. Benchmarks, evaluated using the open-source tool Evalchemy, show OpenThinker2-32B achieving strong results:
- AIME24: 76.7
- AIME25: 58.7
- AMC23: 94.0
- MATH500: 90.8
- GPQA-D: 64.1
- LCBv2: 72.5
These scores indicate its proficiency in advanced problem-solving and logical deduction, often outperforming other models in its class on these specific metrics.
Training & Data
The model was trained for 50 hours across 128 4xA100 nodes. The OpenThoughts2-1M dataset is a key differentiator, incorporating existing datasets like OpenR1 and new, high-quality math and code reasoning data. Further details on the dataset and model development can be found in the OpenThoughts Paper and the OpenThoughts2 blog post.
Intended Uses
OpenThinker2-32B is well-suited for applications requiring robust reasoning, mathematical problem-solving, and code-related tasks. Its Apache 2.0 License allows for broad use.