Chinese artificial intelligence company DeepSeek has unveiled the world’s first open-source AI model capable of achieving gold-medal-level performance at the International Mathematical Olympiad (IMO). The Hangzhou-based start-up announced the launch on 28 November, releasing its Math-V2 model to the public under a permissive licence that allows developers to modify and repurpose the system. The model is available on popular developer platforms Hugging Face and GitHub.
Held annually since 1959, the IMO is regarded as the most prestigious mathematics competition worldwide. Its challenging problems demand deep insight, creativity, and rigorous reasoning, not only from human participants but also from AI systems, according to Harvard University AI researcher Huang Yichen and UCLA computer science professor Yang Lin. Only around 8 per cent of human participants typically achieve a gold medal, highlighting the competition’s difficulty.
Gold-level performance and open access
DeepSeek said its Math-V2 model reached gold-level scores in both this year’s IMO and the 2024 Chinese Mathematical Olympiad. The company claimed the achievement would make advanced mathematical AI tools more accessible to developers globally, coming just months after US AI leaders Google DeepMind and OpenAI reached similar milestones with their proprietary models.
“Imagine owning the brain of one of the best mathematicians in the world for free to explore it for research, fine-tune it, optimise it and run it on your own hardware,” Hugging Face co-founder and CEO Clement Delangue said in a post on X. “No limitations … no company or government to take it back. That’s democratisation of AI and knowledge at its best, literally.”
Self-verifying reasoning and future research
DeepSeek researchers highlighted that the Math-V2 model differs from many existing AI systems, which often prioritise strong performance on benchmark tests at the expense of improving underlying reasoning skills. The company’s model incorporates a “self-verification” process, enabling it to check its answers even for mathematical problems without known solutions.
“This addresses a key bottleneck in current AI systems, which tend to improve only on tasks with easily verifiable solutions,” the researchers wrote in a Hugging Face blog post. They added that the approach could pave the way for more capable mathematical AI systems and have a significant impact on scientific research if further developed.
Unlike DeepSeek, Google DeepMind’s gold medal-winning model is currently available only to subscribers of its premium Ultra plan. OpenAI CEO Sam Altman has stated that the company’s experimental gold medal-winning model will not be publicly accessible for many months.
