Terence Tao and Team Soft Launch the SAIR Open Math OS for Collaboration
A soft launch of the SAIR Open Math OS platform now provides hosted open math models and services for academic teams. This is a very early version released ahead of schedule due to high community demand, and it currently uses donated computer power.
In short: A soft launch of the SAIR Open Math OS platform now provides hosted open math models and services for academic teams. This is a very early version released ahead of schedule due to high community demand, and it currently uses donated computer power.
Mathematicians looking for collaborative tools now have a new shared platform designed to let teams view each other's prompts.
What happened, in plain words
Terence Tao from UCLA announced the soft launch of the SAIR Open Math OS. The platform provides web interfaces and services similar to advanced artificial intelligence models. Right now, it hosts existing open math models called DeepSeek V4.1 Flash, Kimi K3, and GLM 5.3 Flash. These hosted models trail top frontier models in capability by roughly a year or so, meaning they cannot handle one-shot problem solving. However, they can still handle medium-sized Lean formalization, alongside summarizing or discussing mathematical arguments and topics. The platform focuses on team accounts where members share visibility of each other's prompts to aid collaborative projects. Free accounts are currently being granted to people from academic institutions using donated compute, and developers are gathering partners to build a more fine-tuned math model through post-training.
Key points
- Soft launch of SAIR Open Math OS: The platform is now available in a very early version, with developers accelerating their release timeline due to high community demand for open models.
- Current model capabilities: The platform hosts open models like DeepSeek V4.1 Flash, Kimi K3, and GLM 5.3 Flash, which are about a year behind current frontier models and cannot do one-shot problem solving.
- Team-based accounts: Unlike individual accounts, the platform emphasizes team accounts where members can see each other's prompts to better support collaborative projects.
- Access and donated resources: Free accounts are being given to academic requesters using donated compute, though resource allocation may change depending on future demand and donations.
Terms explained
- Lean formalization — The process of translating mathematical ideas and proofs into a precise computer language so a machine can check them. Example: Writing out a geometry proof in a strict computer format so a software checker can verify every single step.
- One-shot problem solving — The ability of a system to solve a complex problem correctly on the very first try without extra hints. Example: Solving a difficult riddle instantly upon hearing it once, rather than needing clues or multiple guesses.
Critical analysis
What is encouraging. The project offers a transparent platform built around open models, relying on academic partnerships and team-based workflows to explore how artificial intelligence can assist collaborative mathematics.
What deserves caution:
- Very early release version. The software is a soft launch released ahead of its original timeline, meaning users should expect it to be a work in progress with features still under development.
- Behind frontier capabilities. The hosted open models lag behind current top models by roughly a year and cannot perform advanced tasks like one-shot problem solving.
- Dependent on donated resources. The platform relies entirely on donated compute and academic requests, which means future availability and account allocations may shift based on changing support.
Bottom line: SAIR Open Math OS is an early-stage, team-focused platform hosting older open math models using donated resources, built primarily to support collaborative academic work.
What we still don't know
The source does not provide performance metrics, exact user counts, specific timelines for future updates, or details on how the upcoming fine-tuned model will be trained beyond general post-training plans.
Based on reporting from Terence Tao (UCLA). This is an independent explainer, written in our own words with AI assistance; Terence Tao (UCLA) has not reviewed or endorsed it. Read the original for the full details.