Microsoft MAI-Base-2 completes pre-training at 2.8T params

Microsoft AI has finished pre-training MAI-Base-2, its next in-house foundation model, the company said on October 15, 2026. The model has 2.8 trillion parameters, with 80 billion active for each token. Microsoft says it beats every open-weight model it tested and took far less computing power than a leading rival. It shows Microsoft building frontier-scale models of its own, alongside the OpenAI models it already offers.

Pre-training is the first and most expensive stage of building a language model. The model reads trillions of words and learns to predict the next one. MAI-Base-2 is now in post-training, the stage that turns a raw model into a useful assistant. It is not available to use yet.

What Microsoft built

MAI-Base-2 is a mixture-of-experts (MoE) model. Only a small part of the network, the "experts" picked for each token, runs at any moment. That is why a 2.8-trillion-parameter model needs only 80 billion parameters of work per token.

Item	MAI-Base-2
Total parameters	2.8 trillion
Active parameters per token	80 billion
Training hardware	About 16,000 Nvidia GB200 GPUs
Training time	38 days, in August and September 2026
Training mix	50 trillion tokens
Image-and-text share	18%, from 4.5 trillion unique tokens
Status	In post-training, no release date

The model learned from images as well as text. Image-and-text data made up 18% of the 50-trillion-token mix, according to Microsoft AI's post. The number of unique training tokens grew by roughly 36% over its predecessor, MAI-Base-1.

How well the training ran

Microsoft shared two measures of efficiency. Model FLOPs utilization, or MFU, averaged 21%. It is the share of the GPUs' peak math power that went into useful training. Goodput was 89%. That is the share of time spent training, as opposed to recovering from crashes and restarts.

Microsoft also took a public position on a sensitive topic. "Our pre-training is built on organic data which does not involve distillation from other models," its post says. Distillation means training a model on another model's answers. This week, Anthropic described new defenses against distillation of Claude.

The efficiency claim

The headline figure is a comparison with DeepSeek-V4-Pro-Base, a leading open model. Microsoft says MAI-Base-2 reaches the same results with 4 to 12 times fewer training FLOPs. The body of the post gives the range as 4.2 to 12.3 times.

"MAI-Base-2 outperforms every open-weight model we tested on our held-out evaluations," Microsoft wrote. Held-out evaluations are tests kept out of the training data. These are Microsoft's own tests, and no outside group has checked them.

Where it fits

Microsoft launched its MAI program publicly on June 2, 2026, at its Build conference, with seven models, Crypto Briefing reports. The team is Microsoft AI, led by Mustafa Suleyman.

The step up in scale is large. MAI-Base-1 was trained on 8,192 GB200 GPUs and 30 trillion tokens of human-written data, plus 3.55 trillion tokens in a later stage. MAI-Thinking-1, built on that base, scored about 52.8% on SWE-Bench Pro, a coding test, Crypto Briefing says.

Microsoft is also building its own AI chips. It plans multi-gigawatt deployments of Maia chips from 2027, according to Crypto Briefing. We covered the Maia 300 chip this week. Work on MAI-Base-3 has already started, Microsoft says.

What this means for developers

There is nothing to call or download today. MAI-Base-2 is a base model in post-training, and Microsoft has given no release date or price.

Watch for two things when it ships. First, check whether the efficiency edge shows up as lower prices on Microsoft's platforms. A model that costs less to train can cost less to serve, but Microsoft has promised nothing. Second, wait for outside benchmarks. "Every open-weight model we tested" depends on which models and which tests.

For teams building on Azure, the bigger point is choice. A Microsoft-trained frontier model would sit next to OpenAI's models on the same platform. If you have tied your code to one model family, keep a thin layer between your app and the model. That way you can compare MAI models against your current choice when they arrive.