Terence Tao says SAIR has soft-launched Open Math OS

On 16 October 2026, Terence Tao posted SAIR Open Math OS released on his blog What’s new. The post is filed under advertising and math.GM. He writes that the service “is now soft launched.” The link on that phrase goes to openmathmodel.org.

The model list in his post is specific. He writes that Open Math OS “offers services and web interfaces similar to frontier AI models, though currently we are hosting existing open math models (DeepSeek V4.1 Flash, Kimi K3, and GLM 5.3 Flash) which are roughly a year or so behind the current frontier models in capability.” He says the hosted models are “not able to perform ‘one-shot problem solving’ of the type we have heard frequently about in recent months,” and that they are “capable enough for many other useful mathematical tasks, such as medium-sized Lean formalization, or summarizing or discussing a mathematical argument or topic.” Those capability lines are Tao’s. This account did not run the models.

Team accounts, compute, and who can get one

Tao writes: “One feature of our model is the emphasis on team accounts rather than individual accounts, in which members of a team see each other’s prompts.” He says that format may suit collaborative projects.

On access, the post says: “We are hosting these models using donated compute, and are in the process of granting free accounts to requesters from academic institutions. We may need to adjust the allocation of compute depending on demand and on future donations, but we will try to make the models available to as many users as possible.”

He also writes: “This is still a very early version of the model; we accelerated our timeline for first release due to the increased demand for open models from the community.” The 16 October post points “a few weeks ago” at his 18 September announcement. That post says he co-founded the Foundation for Science and AI Research (SAIR) with private donors as a nonprofit, and that demand led the group to move an open-model initiative forward.

What the site shows besides the three names

The Open Math Model home page describes “the OS for collaborative math research and scientific computing” and says it is “built on open frontier models, with the mathematical community in control of its data.” The hosted models, as Tao names them, are DeepSeek V4.1 Flash, Kimi K3, and GLM 5.3 Flash.

The same page says every project has a cloud machine that keeps running with the laptop closed, and that more than 100 tools are preinstalled. The graphic names Lean and Mathlib, SageMath, PARI/GP, SymPy, and LaTeX. It describes Ultramath as agents trying different approaches and returning findings to the chat. Those are the site’s product statements.

The page’s data section says projects, files, and chats stay private until someone chooses to share them, and that training on a user’s work requires explicit consent on terms agreed in advance. The apply steps printed on the page are: sign in with a SAIR account, introduce the team and research, and, once approved, invite collaborators. Those three steps do not repeat Tao’s phrase “requesters from academic institutions.” His post is the place that phrase appears.

Partners for a math post-training run

Tao writes that SAIR is “in the process of gathering partners to develop a more fine-tuned model for mathematics by post-training on an existing open model.” The words “this link” go to the initiative page. That page is an invitation from the SAIR Foundation to build open models with the mathematical community. It says the first phase is everyday mathematical work: understanding difficult arguments, checking references, exploring examples, writing code, and formalizing proofs. It says shared models, code, and tools would be released under open licenses, “for example Apache 2.0, MIT, or CC BY 4.0 where appropriate.” It says the community owns the data and that user data would be used for training only with explicit consent. A supporter form asks what kind of help is on offer: funding, compute, data, models and engineering, tools and infrastructure, community, or other. The contact address printed there is [email protected].

Tao also points readers to a Zulip channel for details and feedback. The Zulip server asks visitors to log in.

Medium-sized Lean formalization is one task Tao names. Related reading on this site: a Ghys-conjecture preprint and its Lean repository, and a preprint on OpenAI’s Navier-Stokes Lean files. Tao’s post cites neither.

The service, as Tao describes it on 16 October, is a soft launch of a team chat on donated compute, hosting three named open models, with free accounts for academic requesters still being granted and with partners still being gathered for a mathematics post-training effort. The site behind the link is up. It adds product claims, and two extra model names, that his post does not.