The Hosts' aim is not just to run a competition, but to provide an ecosystem for it.

We thus have partnered with the Fields Model Initiative (supported by two organizations: the LLMC from the National Institute of Informatics in Tokyo and Benchmarks+Baselines from Vienna) and with Thinking Machines, which makes the Tinker service available.

These partnerships provide compute and API credits, respectively, to help contestants train/fine-tune their models. AIMO is the first large-scale Kaggle reasoning competition to provide significant support in this direction, enabling competitors to have a streamlined experience for training and competing with their models.

This addresses a significant pain point from previous AIMO competitions and a persistent issue of any large-scale competitions that focus on computationally expensive AI models, such as LLMs, as previous competitions were 1) skewed towards large labs with significant resources having an edge, and 2) skewed towards teams with strong engineering know-how, putting mathematicians keen on competing at a disadvantage.

Our partnerships solve both problems. AIMO3 levels the playing field by providing access to 128 H100 GPUs via the Fields Model Initiative (applications for more will be considered on a case-by-case basis), totaling hundreds of thousands of hours of compute we can offer to teams, and up to $400 in API credit for Tinker. Tinker abstracts away much of the engineering knowledge, making it particularly beneficial for mathematicians.

Because demand may exceed supply, an application process will be in place for both of these services, and applications will open later in December, with precise details outlined in the Kaggle Discussion forum. Because it is our goal to make these resources available to teams with the greatest promise to devise a strong reasoning model, factors beyond application time may be considered.