China’s DeepSeek Unveils Math Model That Matches IMO Gold Medalists

China-based AI research lab DeepSeek has unveiled DeepSeekMath-V2, an open-weight artificial intelligence model that the company claims demonstrates world-class mathematical reasoning and theorem-proving capabilities.

The model reached gold-level performance at the International Mathematics Olympiad (IMO) 2025), solving five out of six problems—a benchmark typically achieved only by the world’s top high-school mathematicians.

DeepSeekMath-V2 joins advanced Google DeepMind and OpenAI models that also solved 5 of 6 problems earlier this year, becoming the first AI systems to ever reach IMO gold status. Of 630 human contestants worldwide, only 72 earned gold medals.

Beyond IMO, DeepSeekMath-V2 scored top results across other elite competitions, including the China Mathematical Olympiad and the Putnam exam. DeepSeek reported that on Putnam 2024, the model solved 11 of 12 problems completely, achieving 118/120, surpassing the highest human score of 90.

DeepSeek argues that current AI models often excel at producing correct answers but lack the rigorous, step-by-step reasoning required for theorem proving. To address this, the lab developed a dual-system approach: a proof generator paired with a dedicated verifier that grades proofs—not answers. The generator is rewarded only when it identifies and fixes its own reasoning errors.

The Chinese startup continually increases the difficulty of verification to prevent overfitting, creating harder proofs and expanding training data. The open weights of DeepSeekMath-V2 are now available on Hugging Face—a move Delangue called “democratisation of AI and knowledge at its best.”