☰
Third $2.2 million AIMO progress prize launched
Competition created and funded by XTX Markets to spur the development of AI models capable of mathematical reasoning
November 2025
The third progress prize for the Artificial Intelligence Mathematical Olympiad (AIMO3) has launched today on Kaggle.
Building on the success of AIMO1 (won by Project Numina) and AIMO2 (won by Nvidia’s team NemoSkills), AIMO3 increases the prize pot, raises the difficulty level of the math problems, and provides more compute hardware for competitors.
IMO-Level Mathematical Reasoning
AIMO3's math problems are significantly more difficult, ranging from National Olympiad level to International Mathematical Olympiad (IMO) standard. All problems are original, ensuring zero risk of data contamination, and this stage comprises 110 math problems spanning algebra, combinatorics, geometry, and number theory.
Recent breakthroughs by closed-source models achieving gold medal performance at the 2025 IMO demonstrate that AI is approaching human-level capability of mathematical reasoning. Our "OpenAI x AIMO eval" in March 2025 revealed, under more stringent benchmarking conditions, an even more striking result - commercial models solved 47/50 of AIMO2 public leaderboard problems at the first try, with the highest score achieved by a team on Kaggle being 34/50. AIMO3 aims to close this gap between closed- and open-source models.
The AIMO's grand prize of $5m will be awarded to the first publicly-shared AI model capable of winning a gold medal in the IMO in a future stage of the competition.
What Makes AIMO3 Different
Industry-Leading Hardware
Contestants will have access to H100 GPUs to run their models, offering roughly double the compute power of AIMO2. They can now run larger, state-of-the-art open-weight models like GPT-OSS-120B and Qwen3-Next, thanks to the H100's expanded memory and support for advanced formats.
Additionally, most leading models are already optimized for H100s by their providers, allowing contestants to use them directly without the re-optimization required in AIMO2. H100s are available for both training and testing via Kaggle, so the entire workflow stays on the same architecture, maximizing efficiency from development to final submission, and open-sourcing beyond the AIMO.
For select participants (see Kaggle for criteria):
- Up to 128 H100 GPUs are available to fine-tune models before submission through our partnership with the Fields Model Initiative, supported by the LLMC research group from the National Institute of Informatics, Tokyo, and Benchmarks+Baselines, Vienna.
- Tinker credits are available via our partnership with Thinking Machines.
These compute enhancements will enable smaller teams to match the larger labs on a level playing field.
Prizes to reward more than just performance
Beyond the main competition, for the first time, AIMO3 will offer several Extra Prizes (EPs) that replace the Early Sharing Prize from AIMO1 and AIMO2. These include:
- Longest Leader Prize: For the team whose model stays longest on top of the public leaderboard.
- Write-up Prizes: For the best technical explanation of a team's approach.
- MathCorpus Prize: For publishing novel datasets that will help the wider community.
- Hardest Problem Prize: Awarded to the best model that solves the single problem least successfully solved by the other AIMO3 models.
For the full rules of AIMO3 and the EPs, please visit Kaggle's AIMO3 overview and rules pages.
AIMO's Objective
The AIMO Prize exists to accelerate the open development of AI models that can reason mathematically – a true frontier in AI capabilities. By requiring all winning solutions to be publicly shared, we ensure the entire research community benefits from these advances.
With answers also now requiring five digits instead of three, guessing is virtually impossible, forcing models to truly reason through the math problems while keeping numbers manageable for LLMs and understandable to humans.
AIMO2 was one of the largest reasoning competitions ever run, with more than 2,000 teams, and its LLMs have demonstrated state-of-the-art mathematical reasoning among open-source models.
The reward?
Aside from offering the largest prize pot to date, AIMO3 enables participants to join an elite community of model builders and gain valuable learning experience through access to engaged discussion forums. Winners will also showcase their models at AI Day at the 2026 IMO in Shanghai, China - a global stage for the next generation of mathematical AI.
Compete in AIMO3
Whether you’re a researcher at a major lab, a student with a passion for math and AI development, or an independent developer, we encourage you to enter AIMO3.