rewrite this title New AI Model Challenges DeepSeek, Proves Less Can Be More

The open-source model, OpenThinker-32B, achieved similar or better results in key performance tests while requiring far less training data, according to the February 12 blog post.

OpenThinker-32B was trained using just 114,000 examples, a fraction of DeepSeek’s 800,000. The dataset, called OpenThoughts-114k, included detailed solutions, coding test cases, starter code, and subject-specific information.

Training took about 90 hours on four nodes, each equipped with eight H100 GPUs. Another dataset, containing 137,000 unverified samples, was processed separately using Italy’s Leonardo Supercomputer, consuming 11,520 A100 GPU hours in just 30 hours.

When tested, OpenThinker-32B delivered a 90.6% accuracy rate on the MATH500 benchmark, surpassing DeepSeek’s 89.4%. It also scored 61.6 on GPQA-Diamond, compared to DeepSeek’s 57.6, which shows strength in general reasoning tasks.

For coding tasks, OpenThinker-32B lagged slightly behind, scoring 68.9 against DeepSeek’s 71.2. Since OpenThinker-32B is open source, these numbers could improve as developers contribute refinements.

Built on Alibaba’s Qwen2.5-32B-Instruct language model, OpenThinker-32B supports a 16,000-token context window. While this is smaller than other AI models, it is still enough to handle complex equations and long programming tasks.

On February 13, Elon Musk revealed the latest version of xAI’s chatbot, Grok 3. What can it do? Read the full story.