A team of international researchers unveiled OpenThinker-32B, an open-source AI model that achieved a 90.6% accuracy on the MATH500 benchmark, outperforming China's DeepSeek, which scored 89.4%. OpenThinker also excelled in general problem-solving tasks, scoring 61.6 on the GPQA-Diamond benchmark versus DeepSeek's 57.6. Remarkably, OpenThinker accomplished these results using only 114,000 training examples, significantly less than DeepSeek's 800,000. The model, based on Alibaba’s Qwen2.5-32B-Instruct LLM, supports up to 16,000 tokens, suitable for complex tasks but less than current standards. This advancement comes amid increased competition in AI reasoning, particularly with developments from OpenAI and xAI's Grok-3 model. OpenThinker's open-source nature allows enthusiasts to further improve its capabilities, setting it apart from DeepSeek, which keeps its training data private. This release signifies the potential for competitive AI models without relying on vast proprietary datasets, appealing to developers wary of using Chinese models.

Source 🔗