Training LLMs for mathematics that converge to Fields Medal performance. A non-profit initiative.
Our long-term goal is a fully open-source LLM that can genuinely assist mathematicians across all domains, ultimately approaching Fields Medal–level capability. We believe a model of that kind has to be built the way mathematics itself is built: in the open, by many hands, and with results anyone can check with an open design, open data, and an open training pipeline. Mathematicians, scientists, and machine-learning engineers should be able to inspect these models, retrain them, run them on their own hardware, adapt them to their needs, and argue publicly about how they ought to be evaluated and how their outputs should feed into the general mathematical discourse. Our very first model, FM-Pochi-32B — released just before IMO26 to facilitate a clean eval, and presented at AITP26 — came out of a community volunteering effort on a small compute footprint provided by the NII, and demonstrated this is possible. The Fields Model Initiative offers an umbrella under which such development can take place, and we offer occasional compute grants. We support the Leiden Declaration on Artificial Intelligence and Mathematics, and the commitments below are our attempt to put several of its recommendations into practice.
Everything we release carries an open licence: weights, training code, data pipelines, and evaluation harnesses, under terms such as Apache 2.0, MIT, or CC BY 4.0. We document how a model was trained and where its training data came from, including the permissions attached to that data and any restrictions on redistributing it. If you can contribute expertise, compute, or other support, please get in touch. We work organically, and once a good team has come together, we can get to work. The initiative is led by Simon Frieder, former manager of the AIMO Prize, a competition in which thousands of teams took part and which produced open LLMs.
Another important aspect of our work is community engagement. We are happy to curate community-provided models, evaluate them and certify they are useful to mathematicians in some form. Mathematicians should have vetted tools they use.
While FM-Pochi-32B was the beginning and showed what is possible, we are keen to move to research-level mathematics, and to move the needle in terms of fully open models. We focus not only on models but on the entire ecosystem around them. Currently, the state of mathematical benchmarks and datasets is skewed towards specific types of datasets (proof generation), and we welcome contributions from mathematicians in terms of data and evaluations, and from engineers in terms of training know-how.
The initiative began as an informal effort shaped by early intellectual exchange between Simon Frieder and Pontus Stenetorp. Simon provided direction and continuity, gave the initiative its name and broader identity, and is now working to incorporate it as a non-profit. Pontus and his group at NII contributed engineering support and compute grants. Philip Vonderlind and Michal Stefanikjoined during the initiative's first partnership with the AIMO Proof Pilot, supporting contestants with compute to train their models.
That collaboration produced a fully open-source model that beat a version of Opus 4.8 on IMO-level problems. Building on this work, the initiative is hosting Fields Model Day to bring the wider community together around AI4Math projects.
AIMO Proof Pilot focused on assessing whether fully open LLMs (the Olmo family, the SmolLM family, and Apertus), each of which has detailed training mixes information and checkpoints released, can be adapted and posttrained to output proof at IMO level. The competition ran for one month, with invited teams competing, and the final eval took place on June 25th, 2026. After the models were uploaded to Kaggle, six IMO-level problems were served and their proof outputs returned. These were graded by AIMO on a markscheme published in advance by a team of graders with IMO experience, with at least two graders per proof. Each competing team from the AIMO Proof Pilot was given access to up to 24 H200 GPUs through LLMC, NII to train their models, as well as engineering assistance, throughout May 2026 to mid June 2026. The data fee was waived to allow contestants to fully focus on creating the best-possible models.
The AIMO Interpretability Challenge is a competition accepted at NeurIPS 2026 that focuses on distinguishing robust from spurious reasoning in LLMs, as a strong final-answer accuracy may not reveal whether a model relies on stable reasoning mechanisms or exploits brittle reasoning shortcuts. Competitors need to upload a system that assesses whether a given model solves a given problem reliably. Building on AI Mathematical Olympiad (AIMO) problems and submissions, together with resources from the Fields Model Initiative, the competition will provide (1) newly-published olympiad-level math reasoning problems and their symbolic representations, allowing generation of novel functional variants, (2) access to frontier reasoning models, and (3) assessments of models' adversarial robustness on these problems. Contestants are supported by compute from the LLMC, NII through the Fields Model Initiative.
Pontus Stenetorp's group and LLMC, NII provide access to compute resources and related engineering support for grant recipients.
To apply for a compute grant or discuss a project, email us at hello@fieldsmodel.org.