Mistral Large 4 Joins Open Math OS — and the Open Part Comes With a Waiting List

On 3 November 2026, Mistral Large 4 became the newest model on Open Math OS, the SAIR Foundation's shared workspace where mathematicians and AI agents work on the same projects. The announcement was a short post rather than a launch page: 1.05T open weights, a 1M context window, live that day for approved teams. That last clause is the one worth sitting with. Open Math OS is built only on open models precisely so the mathematical community can inspect and run what it is relying on, and the newest addition still arrives behind an approval queue. So do the three models already on the roster — DeepSeek V4.1 Flash, Kimi K3 and GLM-5.3-Flash.

What Mistral actually said, and what it left out

The post carries two technical claims and one access term. The claims — 1.05 trillion total parameters and a one-million-token context window — match the checkpoint Mistral published on 29 October, so they are not marketing arithmetic. The access term is "approved teams," with no description of what approval requires, how long it takes, or whether an individual researcher can get in without a team.

What the post does not contain is just as informative. There is no mathematics benchmark. There is no statement that Mistral Large 4 was evaluated on Lean proofs, on formalisation, or on any of the tasks Open Math OS was built for. There is no note on whether the model gets the same tool access as the incumbents — the shell, the PDF tools, the Lean checker, the MCP servers — or a reduced set. And there is no pricing, which is consistent with a workspace that runs on donated compute rather than metered tokens.

None of that is a criticism of the model. It is a description of how much is still unknown about the pairing, six days after the weights landed.

Open Math OS is a workspace, not a model

The distinction matters because the name reads like one. Open Math OS is a SAIR Foundation initiative, soft-launched on 16 October 2026, and Terence Tao — a SAIR co-founder — described it as a hosted service with "services and web interfaces similar to frontier AI models" rather than anything you install. Its open-source runtime, the Open Math Harness, is written in Rust and structured as a capability-based microkernel: every program, agents included, runs as a confined process, and the stated design principle is "agents propose, mathematicians decide."

Around that kernel sit the things that make it a workspace rather than a chat window. Team accounts, where members can see each other's prompts. Persistent project memory. A parallel-exploration mode called Ultramath that runs several agents against the same problem at once. Lean-verified proof search. Support for MCP servers and for plugins that users build themselves. SAIR is granting free accounts to academic requesters and running the whole thing on donated compute, which is why account caps and allocation adjustments are explicitly on the table.

The roster it launched with was entirely Chinese open-weight models. That is the gap Mistral Large 4 fills: it is the first model on Open Math OS from a European lab, and the first with a French sovereign-deployment story behind it. For a foundation whose pitch is community control of data, having a non-Chinese, non-American option is a real addition rather than a checkbox.

The licence question nobody in the thread asked

Mistral Large 4 is not Apache 2.0, and that is a change from its predecessor. Its weights ship under the Mistral Ecosystem Open-Weight License, a modified MIT licence that allows use, modification, fine-tuning and commercialisation — with one carve-out. Anyone offering the model as a model-as-a-service must obtain a separate commercial licence from Mistral.

Open Math OS is a hosted service that lets teams call a model over a web interface. Whether that arrangement is "model-as-a-service" in the licence's sense, or falls under the exemption for embedded end-user products, or fits the narrow carve-out for merely relaying requests to models hosted elsewhere, is a legal question SAIR has not addressed publicly. Mistral has not said either. It is entirely possible — likely, even — that there is a commercial agreement behind the announcement that neither party has described. But a reader who saw "open weights" and assumed the hosting question was settled should know that it is not, because the same clause will apply to every lab that wants to build a shared workspace on this model.

Why a 1M window is more than a spec-sheet line here

Context windows are usually sold on document volume, and a million tokens sounds like a filing cabinet. In a formal-mathematics workspace the useful version is different. A project accumulates a Lean file, its imports, a proof state, the discussion around a failed tactic, and the agent's own session history. All of that has to be in the same frame for an agent to do anything but guess. Threads that survive across days are the point of a shared workspace; a window that truncates them turns every session into a cold start.

That is an argument for why Mistral Large 4 fits the design, not evidence that it performs. The two are easy to confuse in a launch week, and Mistral's post does not help — it lists the window as a headline number next to the parameter count, which is spec-sheet grammar, not an evaluation.

What an approved team gets, and what is still unstated

• Evaluation — no published mathematics results for Mistral Large 4 on Open Math OS; the roster's launch models were described by Tao as roughly a year behind the frontier and good at medium-sized Lean formalisation, which is a capability claim, not a score.

• Tooling parity — unstated whether the new model gets Ultramath, proof search and the full scientific toolchain, or a limited subset while it is being assessed.

• Limits — no published rate limits, quotas or context caps inside the workspace, where a 1M window may be trimmed by the platform regardless of what the model supports.

• Data policy — Tao wrote in October that SAIR wanted to "first collect some usage data before finalizing our policies." That is an honest statement, and also a reason to read the current terms before putting an unpublished proof into a project.

• Selection — no explanation of how a team picks a model per project, whether the choice is per chat or per project, or whether SAIR routes between them automatically.

The pattern is consistent: the announcement describes a model being added to a lineup, not a capability being demonstrated. If you are deciding whether to apply, that is the honest state of the evidence.

If what you want is the models rather than the workspace

Open Math OS is the interesting part of this story, but it is not the only way to reach the models it runs. GLM-5.3-Flash is in the OrcaRouter catalogue at provider list price, and so is Kimi K3 — two of the three models the workspace launched with, callable from one API alongside more than 200 others, with automatic failover between providers and 0% markup on the provider's own rate. The routing DSL will compose them into a single call if you want a panel rather than a single answer, which is a rough approximation of what Ultramath does inside the workspace without the Lean checker attached.

Mistral Large 4 is not in our catalogue, and we are not going to pretend otherwise. If you want that model specifically, the routes today are Mistral's own API, the open-weight checkpoints on Hugging Face, or the vendor's third-party hosting partners. The trade-off is real: the workspace gives you the shared project, the tools and the proof checking; the API gives you the model alone, at a price you can measure.

What to watch next

Three things would turn this from a roster update into a result. A per-model evaluation inside Open Math OS — even a small one, a set of Lean formalisations scored the same way for all four models — would be the first public evidence about which model a mathematician should pick for which job. A published access policy would resolve the tension between "open" and "approved." And a statement from SAIR or Mistral on how the hosting arrangement sits against the model-as-a-service clause would tell every other lab whether they can build the same thing.

Until then, the accurate summary is narrow: a trillion-parameter open-weight model with a 1M window is now one of the options in a research workspace that is still, by its founder's own description, a very early version.