We are excited to deepen our commitment to open source by becoming a founding partner of Project Numina, a non-profit organization dedicated to advancing frontier AI in mathematics. In under 6 months, Numina has assembled a world-class team of contributors, collected and curated the largest mathematics competition level problem-solution dataset in history, and trained a frontier AI model that won first prize at the inaugural 2024 Artificial Intelligence Mathematical Olympiad. We believe Numina’s achievement reinforces the virtuous open source flywheel championed by Mistral AI and HuggingFace, and we are energized to help catalyze this Europe-led movement that is accelerating solutions to the world's most pressing scientific challenges. 

Mathematics is the bedrock of science and human progress, serving as a foundation for many technological advancements. From Euclid to Newton, Fourier and Ramanujan, mathematicians have discovered patterns, structures, and reasoning frameworks that have fueled human ingenuity and unlocked our greatest potential. That’s why we have partnered with several leading companies improving both human and artificial intelligence in mathematics, including Multiverse, Numerade, Symbolica, and Mistral.

When Yann Fleureau, Guillaume Lample, Stanislas Polu, Jia Li and Helene Evain founded Numina with the mission of advancing frontier AI in mathematics, we saw a non-profit organization committed to forging new horizons for today’s scientific Magellans through the principles of open source. We believe Numina’s founding team not only has the unique talent and expertise that sets them apart to take on this challenge, they are also fiercely ambitious, curious and purpose-driven – the type of people we dream of partnering with.

With an initial focus on winning the Artificial Intelligence Mathematical Olympiad (AIMO), the team and their community created the largest and highest quality public dataset in AI4Maths in history with 860,000 pairs of competition level problems and solutions. Then, they trained Numina-Math-7B, a new state-of-the-art model that achieved top spot in the competition with a score 29/50. A performance and winning margin that Field’s Medal winner Terence Tao himself admitted was “higher than expected” (see demo here).

Now, Numina has made the full inference and training code, model weights and dataset publicly available under an Apache 2.0 license (see here). By inviting the research community to join their open scientific collaboration, Numina hopes to foster open standards and infrastructure for AI reasoning and accelerate progress on a national, regional and global scale.

As fervent supporters of open source and mathophiles, it is a privilege to partner with Numina. We congratulate the team on their remarkable achievement, and invite everyone to contribute to their mission to foster the development of human and artificial intelligence in mathematics.