The Pulse
Terence Tao’s SAIR Opens Plan for Open Math Models
The Foundation for Science and AI Research has announced an initiative to build open models for mathematics, inviting partners to contribute funding, compute, expertise, or community-building support.
AI.info Team ·
“We had initially planned a gradual rollout of this initiative, starting with a few pilot projects over the next few months, but given current events and the high demand for open models, we are accelerating our schedule and moving our initial announcement of the initiative to today, despite the fact that some key planning is still underway.”
Terence Tao, SAIR co-founder
The Foundation for Science and AI Research has opened a new effort to build open models for mathematics, with Terence Tao saying the organization is moving faster than originally planned because demand for open systems has increased. In a post dated September 18, 2026, Tao said SAIR was inviting expressions of interest from potential partners.
The announcement does not include a model release, a named industry partner, a training budget, or a delivery date. It describes an early-stage program intended to give mathematicians access to models they can inspect, modify, reproduce, and run outside a single commercial platform.
SAIR’s First Target Is Everyday Mathematical Work
The initiative’s first phase will focus on tasks that sit between literature review, experimentation, programming, and formal proof. SAIR lists understanding difficult arguments, checking references, exploring examples, writing code, and formalizing proofs as initial use cases.
That emphasis separates the project from a contest aimed only at solving olympiad problems or maximizing a benchmark score. SAIR says it plans to judge the systems by the reliability of their assistance, the verifiability of their results, and the cost of sustained use. The foundation has not published a target score or a public evaluation set for the proposed models.
Researchers will be asked to help determine how the systems are trained, what they are tested on, and how they should serve research and education. SAIR says participation should remain open to institutions, regions, and career stages rather than being limited to a small group of laboratories.
Open Weights, Published Methods and Documented Data
SAIR’s proposal calls for open-licensed model weights and code, published training methods, and reproducible evaluations. It also says training data should come with documented sources and compatible permissions, with restrictions on redistribution stated clearly.
The foundation says future releases will report failures and limitations alongside successful results. Independent teams should be able to reproduce the work and adapt the systems to their own mathematical needs, although SAIR has not yet specified which licenses will apply to particular models or datasets.
The proposal names Apache 2.0, MIT, and CC BY 4.0 as examples of licenses that could be used where appropriate. It also says contributors should agree on ownership, licensing, and attribution for jointly developed work from the beginning, while retaining rights to their independent and prior research.
Researchers Would Control Use of Their Data
Data governance is a central part of the plan. SAIR says the mathematical community should decide how shared data is used and that users’ data will be used to train or improve models only with explicit consent and previously agreed terms.
The foundation also promises public governance rules and public decisions. Members should have ways to propose changes, participate in decisions, and hold the initiative’s leaders accountable.
That structure is still a proposal rather than an operating system for the project. SAIR has not published a governing board, voting procedure, data-review process, or partner agreement alongside the announcement.
Partners and Affordable Compute Come Next
In the same post on his website, Tao says SAIR had been negotiating with academic and industry partners on open-weight models for science and open-source tools for using both open and closed language models.
SAIR is seeking organizations and individuals able to contribute funding, compute, technical expertise, or community-building support. Tao says the organization expects to provide more information about its partners and future plans, including how the effort will connect with SAIR’s existing podcasts, events, and mathematical competitions.
The foundation says industry partnerships must preserve the research independence of the mathematical community. It also presents affordable compute as a shared public resource, but gives no estimate of the hardware, training scale, or operating cost required to build the proposed systems.
A Public Commitment Before a Model Release
SAIR’s announcement is therefore a statement of direction and an invitation to collaborate, not evidence that an open mathematical model is ready for use. The organization has set out principles for licensing, data consent, evaluation, and governance while leaving the technical program to be defined with future partners.
The immediate test is whether SAIR can turn those principles into a funded project with public methods, identifiable contributors, and a reproducible release. For now, the concrete action is the partner call, with detailed announcements still to come.