OpenThinker-32B achieves groundbreaking results with only 14% of the data required by DeepSeek. It’s a win for open-source AI.
A team of international researchers from leading academic institutions and tech companies stirred the AI sector Wednesday with a new model. It rivals DeepSeek and often performs even better. The results were reported by DeCrypt.
OpenThinker-32B, developed by the Open Thoughts consortium, achieved an accuracy score of 90.6% on the MATH500 benchmark, surpassing DeepSeek’s 89.4%. The model also performed better on general problem-solving tasks, scoring 61.6 on the GPQA-Diamond benchmark. There, DeepSeek’s Qwen 32B distillation achieved 57.6. On the LCBv2 benchmark, it achieved a solid 68.9, demonstrating strong performance in various test scenarios.
In other words, the model outperformed an equally sized version of DeepSeek’s distillation on general scientific knowledge (GPQA-Diamond). It also beat DeepSeek on MATH500, although it underperformed on the AIME benchmark. Both benchmarks measure math skills.
The power of open source
On programming, OpenThinker scored slightly lower than its Chinese competitor, with 68.9 points versus 71.2. But because the model is open-source, these scores could improve significantly once the community refines it.
What makes this performance unique is its efficiency. OpenThinker needed only 114,000 training examples to achieve these results, while DeepSeek used 800,000.
Detailed metadata for each problem
The OpenThoughts-114k dataset contains detailed metadata for each problem. Namely thorough solutions, test cases for code problems, starter code where needed and domain-specific information. The custom Curator framework validated code solutions using test cases, while an AI judge handled mathematical verification.
A Chinese AI lab not only built a cheaper AI model with DeepSeek. It also exposed the inefficiencies of the entire industry. DeepSeek’s breakthrough showed how a small team, with cost savings in mind, could develop AI models in a whole new way. While tech giants such as OpenAI and Anthropic spend billions of dollars on computing power, DeepSeek would achieve similar results for just over $5 million.
DeepSeek gave sector a boost
The AI sector got a boost after DeepSeek’s demonstrated performance similar to OpenAI’s GPT-4o, but at significantly lower cost. DeepSeek R1 is free to download, use and modify, and its training techniques have also been made public. But unlike Open Thoughts, which made everything open-source, the Chinese development team kept its training data secret.
This means developers are likely to understand OpenThinker better and reproduce it more easily than DeepSeek because they have access to all the pieces of the puzzle.
Reliable alternative
For the broader AI community, this release proves once again that it is possible to build competitive models without huge, proprietary datasets. Moreover, OpenThinker could be a more reliable alternative for Western developers who are still hesitant about using a Chinese model, even if it is open-source. OpenThinker is available for download at Hugging Face. A smaller, less powerful version with 7 billion parameters is also available for less powerful devices.
The Open Thoughts team brought together researchers from several U.S. universities, including Stanford, Berkeley and UCLA, as well as the Juelich Supercomputing Center in Germany. The U.S.-based Toyota Research Institute and other European AI institutions are also supporting the project.