DeepSeek’s math model hits Olympiad gold and goes fully open-source

Deepseek

DeepSeek, the artificial intelligence start up based in Hangzhou, has become the first company to release an open-source AI model that reaches gold medal level performance in the International Mathematical Olympiad. The model, called Math V2, is now available for anyone to download on Hugging Face and GitHub under a permissive licence that allows modification, repurposing and local deployment.

The IMO has been held every year since 1959 and is widely regarded as the most demanding mathematics competition in the world. It is also considered a benchmark for AI systems because the problems require deep insight, creativity and the ability to show reasoning rather than offer short answers. Only around 8 per cent of human participants achieve a gold medal.

DeepSeek’s model has delivered gold level scores not only on this year’s IMO questions but also on the 2024 Chinese Mathematical Olympiad. The release is expected to lower the barrier for researchers and developers globally, especially at a time when Google DeepMind and OpenAI have kept their own high scoring maths models firmly behind premium or experimental access.

Clement Delangue, co founder and CEO of Hugging Face, described the release as a step towards genuine democratisation of AI. He said users could now work with “the brain of one of the best mathematicians in the world” for free, without the risk of a company or government pulling it back.

In a blog post, DeepSeek researchers said many existing systems excel at beating maths benchmarks without improving true reasoning. Instead of chasing scores, Math V2 has been trained to “self verify” its own answers, including for problems with no established solutions. The team argues that this approach removes a major bottleneck in current AI development, where models mainly improve on tasks with straightforward answer checking.

The researchers say there is much more to do but believe self verifiable mathematical reasoning could help future AI systems become significantly more capable.