The Foundation for Science and AI Research, a nonprofit Fields Medalist Terence Tao co-founded with private donors, is opening its bid to build open weight AI models for mathematics research. Tao announced the initiative, called SAIR Open Models, on his blog on September 18, moving up a rollout the foundation had originally planned to stage gradually through pilot projects.
SAIR was set up to support AI uses in mathematics and other sciences independent of the major AI labs. Until now its work has mostly meant podcasts, small events, and competitions, funded in part by a recent grant from trading firm XTX Markets. The open models effort marks a jump in ambition: Tao says the foundation has been negotiating with academic and industry partners to build both open weight models and open source tooling for using AI, open or closed, in scientific work.
Tao frames the accelerated timing as a response to “current events and the high demand for open models,” though the blog post does not name a specific trigger. Some of the initiative’s planning is still unfinished, according to the post, and sign-up for expressions of interest is now open on the announcement page. SAIR is asking prospective partners for four things: money, compute time, subject-matter expertise, or help organizing the community around the project.
The pitch to researchers centers on control mathematicians rarely get with commercial AI tools. Under the plan, the mathematical community, not a company, decides how the models are trained, what they are evaluated against, and how they serve research and teaching. SAIR says training data will carry documented sources and explicit permissions, and that model releases will disclose failure modes alongside results. Outside groups, per the post, will get the ability to redo the training process on their own and reshape the resulting models for whatever their specific research demands.
Ownership questions get more detail than most model announcements bother with. SAIR states plainly that the mathematical community owns the data, and that any use of a researcher’s data to train or improve a model requires explicit consent under terms agreed in advance. Outputs built through the initiative, meaning models, code, and tools, would ship under open licenses such as Apache 2.0, MIT, or CC BY 4.0. Researchers keep ownership of prior or independent work they choose to contribute, and credit is meant to attach to individual mathematical contributions rather than disappear into a shared corpus.
Governance is designed to sit with the community rather than SAIR itself. The foundation says it will work with industry partners for compute and other resources, but that those partnerships need to preserve the community’s research independence, with governance rules and decisions made public and open to challenge from members.
The first phase covers ordinary research chores: writing code, formalizing proofs, exploring examples, checking references, and working through arguments that are hard to follow. SAIR frames success in three parts: how reliably the tool assists, how verifiable its results are, and what it costs to run over time. That treats accuracy and affordability as coequal metrics rather than leaning on benchmark scores alone.
This is a foundation-level announcement, not a shipped model. SAIR has not disclosed which academic or industry partners it is negotiating with, what compute commitments are in place, or a timeline for a first release, and the post itself acknowledges key planning is still underway. Whether an open, community-governed alternative to frontier lab tooling can match the resourcing of a well-funded commercial model will depend entirely on which industry partners actually sign on, and on what compute terms.
For mathematics departments and research groups already leaning on commercial AI tools for proof-checking and literature review, SAIR’s call for expressions of interest is the first concrete opening to shape a model built to their own evaluation standards rather than a vendor’s benchmark suite.
Reported from Terence Tao’s blog, published September 18, 2026.
