A Chinese social app’s AI model, developed by RedNote and named dots-note-3.0, achieved a perfect score of 42 at the International Mathematical Olympiad (IMO), where detailed written proofs are evaluated line by line by human graders. This model, which is still in beta, stands out as only seven out of 666 human contestants matched such a score in Shanghai this year. The model’s innovative approach involved reading the original contest documents directly and using a unique agentic loop that combined natural-language reasoning with self-executed Python code, allowing it to iteratively test and refine its mathematical solutions before final submission.

RedNote: RedNote is the company that operates Xiaohongshu, a leading Chinese social media and lifestyle platform. It developed the dots-note-3.0 AI model, still in beta, which achieved a perfect score on the International Mathematical Olympiad by directly processing contest documents through an agentic reasoning loop. The company has stated it intends to open source the model at a later date.
Rohan Paul: Rohan Paul is a theoretical physicist who publicly commented on the RedNote AI achievement. He noted the significance of the perfect IMO score in relation to the model’s demonstrated reasoning capabilities. His perspective adds expert context to the breakthrough in AI mathematical problem-solving.
Xiaohongshu: Xiaohongshu is a prominent Chinese social app focused on user-generated content around lifestyle, shopping, and personal experiences. As the core product of RedNote, it provided the foundation for the company’s AI research that produced the dots-note-3.0 system achieving a perfect IMO result. The app’s resources and user base have supported RedNote’s broader development of specialized AI models.

Model Family: The dots-note-3.0 variant belongs to the larger dots3 family, which includes other versions optimized for different computational scales and application needs.
IMO Evaluation: The International Mathematical Olympiad requires detailed written proofs that are assessed line by line by human graders, with points deducted for skipped cases or unstated assumptions even when the final answer is correct.
AI Reasoning Approach: Agentic systems that combine natural-language reasoning with self-executed code allow models to test, identify errors, and iteratively repair their own mathematical solutions before submission.