MAI-Base-2: Microsoft AI's 2.8T In-House Base Model, Explained

What Microsoft AI says it finished

We now have state-of-the-art pre-trained base models in the 3T-parameter weight class, built entirely in-house at MAI.

On October 15, 2026, Mustafa Suleyman said Microsoft AI (MAI) had finished state-of-the-art pre-trained base models in the 3-trillion-parameter weight class, built entirely in-house. The model in that announcement is MAI-Base-2: 2.8 trillion total parameters, with 80 billion activated. That split is the spec that matters. The full weight count places it in the 3T class; the activated count is the slice that runs when the model is used.

He pointed readers to a technical write-up for the details and said results from the next training stage were still coming. MAI-Base-2, as described so far, is a finished pre-train, not a shipped assistant. The claim of being state of the art applies to these pre-trained base models in that weight class.

What the frontier pre-training write-up covers

The official page lives on microsoft.ai under the title Frontier Pre-training at Microsoft AI. Its MAI-Base-2 material is organized around comparison, scaling, and the machines the model was trained on. Headings name the arguments: how MAI-Base-2 compares, robust scaling ladders, isolating modeling efficiency from compute, scaling laws for hyperparameters, and codesigning infrastructure with MAI-Base-2.

Those headings tell you the kind of case Microsoft AI is making. It is a pre-training story about scale, tuning, and infrastructure design, not a product spec sheet. The announcement summary does not reprint score tables, rival model names, token budgets, or hardware counts, so those sections should be read on microsoft.ai rather than treated as a published leaderboard.

How it compares, and what is still open

The only head-to-head claim in the announcement itself is qualitative: state of the art among pre-trained base models in the 3-trillion-parameter weight class, from a lab that built the model in-house. MAI-Base-2 is a base checkpoint. Any later assistant, coding agent, or reasoning model would sit on top of further training, and Suleyman said those results are still to come. Until that lands, there is no fair way to rank it against finished chat or agent models that have already been tuned.

Nothing in the announcement lists an app, an API, a price, a rate limit, or a weight download. The page you can open today is the technical write-up on microsoft.ai. Open questions are the ones the schedule already flags: what the next training stage changes, whether Microsoft AI will expose the base model or only later derivatives, and how the comparison section's evidence reads once you are on the page.

• A base model, not a product

  Pre-training is complete. The announcement does not describe a consumer app or a developer endpoint you can call.

• Next-stage results still ahead

  Suleyman said those results are coming and asked readers to watch for them. Quality after that stage is unpublished.

• Read the comparison on the page

  How MAI-Base-2 compares is a heading in the write-up. Numeric bake-offs are not restated in the announcement summary.