The SAIR Foundation has soft-launched Open Math OS, a web workspace for collaborative mathematics research. The October 16 release brings team projects, hosted AI models and proof-checking software into one environment. Terence Tao’s launch post says it initially uses existing models, rather than a new mathematical model trained by SAIR.
Natalia Krizhanovskaya writes formulas on a blackboard in Petrozavodsk, Russia. The 2021 photograph is illustrative and does not depict or show an affiliation with Open Math OS. Photo by Andrew Krizhanovsky, released to the public domain.
Tao said the hosted models could help with medium-sized proofs written in Lean, a programming language and theorem prover that lets mathematicians express proofs for machine checking. He said they were roughly a year behind current frontier models in mathematical capability and could discuss or summarize arguments, but were not yet suited to one-shot problem solving.
Open Math OS gives each project a persistent cloud machine with Lean, Mathlib and more than 100 mathematics and scientific-computing tools, according to the project site. A successful Lean check confirms that a written proof follows from its formal statement; it does not show that the statement matches the mathematician’s intended question.
The public model selector lists five options: the three models named in Tao’s October 16 launch post—Kimi K3, DeepSeek V4.1 Flash and GLM 5.3 Flash—plus GLM 5.3 and Nemotron 3. The site’s About page still names only the original three, and neither page dates the other two entries or says whether all five are available to approved teams.
Tao says team members can see one another’s prompts. The product site describes chats, files and projects as private until shared, while also saying collaborators share projects and conversations; it does not specify the visibility boundary for prompts within a team. The site says work will be used for model training only with explicit consent.
Access is still being opened gradually. Tao said free accounts for academic-institution requesters were being granted on donated computing capacity, with allocations subject to change as demand and donations shift. The site’s application flow asks teams to describe their research and receive approval before inviting collaborators.
SAIR is still gathering partners to post-train an existing open model for mathematics. That remains a future step: Tao’s October 16 post gave no release date for an SAIR-trained model.
Image credit: Andrew Krizhanovsky, Natalia Krizhanovskaya writing at a blackboard, Wikimedia Commons. Public-domain dedication by the author; illustrative image, unrelated to SAIR or Open Math OS.
