Dashboard

Microsoft MAI-Base-2: 2.8T parameters, no release date

Microsoft finished pre-training a 2.8 trillion parameter model and published the numbers. Separating what it measured from what it projected.

Cecilia Iona
Cecilia Iona
Senior Editor, AI & Product
17 October 20261 min read

Microsoft AI said on 15 October 2026 that it has finished pre-training MAI-Base-2, a mixture-of-experts model with 2.8 trillion total parameters and 80 billion active. You cannot use it. The model is still in post-training, there is no release date, no price, and no public endpoint. Microsoft says it plans to license the weights to trusted partners through its Foundry platform rather than publish them.

That gap between announcement and availability is the whole story for anyone building on these models. Here is what the Lab Notes post actually claims, and which parts are measurements rather than projections.

What Microsoft measured

The training run itself is reported in concrete terms. These are the numbers with the least room for interpretation:

Figure

Value

Total parameters

2.8 trillion

Active parameters

80 billion

Hardware

16,000 GB200 GPUs

Duration

38 days, August to September 2026

Training mixture

50 trillion tokens

Vision-language share

18 percent of the mixture, 4.5T unique tokens

Average MFU

21 percent

Goodput

89 percent

Two of those deserve a note. MFU, model FLOPs utilisation, is normalised against the GB200's theoretical BF16 peak of 2.5 PFLOPS and excludes rematerialisation and optimiser FLOPs, so it is not comparable to an MFU number computed a different way. Goodput at 89 percent means roughly one day in nine produced no training progress, which is ordinary for a run at this scale and worth knowing if you ever budget your own.

Microsoft also says unique training tokens grew about 36 percent over MAI-Base-1, that the data is organic rather than distilled from other models, and that it processed every training token in-house. The model is pre-trained natively with vision in an encoder-free architecture, so vision is not a later bolt-on.

What Microsoft estimated

The headline efficiency claim is a different kind of number. Microsoft says MAI-Base-2 is 4.2 to 12.3 times more pre-training FLOPs-efficient than DeepSeek-V4-Pro-Base, and 6.3 to 12.7 times better than MAI-Base-1. Both ranges come from fitted scaling laws, not from running the two models head to head. A range that wide, roughly a factor of three between its own ends, is a signal of how much the fit is doing.

The comparison set is also uneven. DeepSeek-V4-Pro-Base is a base model. Kimi K3, Qwen3.8-2.4T-A95B and MiMo-V2.6-Pro-RL were only available as post-trained checkpoints, and Microsoft notes that post-training may affect those results. Evaluation uses bits-per-byte on held-out text, which is tokenizer-agnostic and a reasonable choice, plus held-out code from Microsoft's own codebase and commissioned STEM, maths and question-answering sets.

None of that makes the claim wrong. It makes it a projection from Microsoft's own fits, measured on Microsoft's own held-out sets, and that is a different thing from a benchmark you could reproduce.

Why a pre-training announcement matters to builders

A base model is not a product. Pre-training produces something that continues text well and follows no instructions. Post-training is what turns it into a model that answers questions, uses tools and refuses things, and it is where most of the behaviour you would actually build against gets decided. Microsoft announcing the end of pre-training tells you a capability ceiling may be coming, not that anything has shipped.

The practical reading is narrow. If you are choosing a model this quarter, MAI-Base-2 is not a candidate. If you are choosing a platform for next year, the Foundry licensing route signals Microsoft intends to sell frontier weights to enterprises rather than compete on open weights, which is the opposite of what Reflection released with Beam a day later. Those are two different bets about who gets access to frontier capability.

Microsoft adds that preparations for MAI-Base-3 are already underway, with no specifications or dates. Treat that as a roadmap signal and nothing more.

Common questions

Can I access MAI-Base-2 today?

No. It is in post-training with no release date, no pricing and no public API. Microsoft says it will license weights to trusted partners through Foundry.

What does 80 billion active parameters mean?

In a mixture-of-experts model, only a fraction of the network runs for any given token. MAI-Base-2 holds 2.8 trillion parameters in total but uses about 80 billion per token, which is what sets the inference cost.

Is MAI-Base-2 better than DeepSeek-V4-Pro-Base?

Microsoft's own held-out evaluations say yes on pre-training efficiency, via scaling-law estimates rather than direct measurement. No independent evaluation exists, because the model is not available to evaluate.

Should I plan around it?

Only loosely. Planning around an unreleased model is mostly a matter of not committing to capabilities you cannot test. The announcement changes nothing you can build this month.

For context on the two stages being discussed here, see what pre-training does and what post-training adds. If you want to track releases like this as they land, our guide to following AI news covers which sources are worth the time.

How did this land?

About the author

Cecilia Iona
Cecilia Iona

Senior Editor, AI & Product

Cecilia leads the Swarmz editorial desk. She has spent a decade turning complex AI and product topics into writing people actually finish, and she owns the blog's quality bar.

Share

Get the next post in your inbox

One email a month. Product updates, engineering posts, and the best of Built with Swarmz.

I agree to receive emails about AI building tips and Swarmz product news. Unsubscribe any time.