Microsoft AI, the company’s in-house model lab, says it has finished the first major stage of training a new foundation model, MAI-Base-2. The model is not released, Microsoft has given no date for when it will be available, and all of the performance comparisons so far come from Microsoft’s own tests.
The lab announced the milestone in a Microsoft AI blog post on 15 October 2026. “We’ve recently completed pre-training our new state-of-the-art model, MAI-Base-2, an MoE foundation model with 80 billion active and 2.8 trillion total parameters,” it said.
What MAI-Base-2 is
Pre-training is the stage where a model learns from a huge body of text and other data. Its output is a “base model,” which still needs further training, called post-training, before it becomes a usable assistant. Microsoft says “MAI-Base-2 is currently undergoing post-training.”
MAI-Base-2 is a mixture-of-experts (MoE) model. Only a fraction of it, 80 billion parameters, is used for each token, out of 2.8 trillion in total. That design lets very large models run with less computing power.
It follows MAI-Base-1, on which Microsoft built MAI-Thinking-1, a reasoning model announced in June.
Microsoft’s own performance claims
Microsoft says MAI-Base-2 “outperforms every open-weight model we tested on our held-out evaluations.” Held-out evaluations are tests the lab keeps private. Microsoft has not published the full list of models it compared against.
It also says MAI-Base-2 is “4-12x more pre-training FLOPs-efficient” than DeepSeek-V4-Pro-Base, which it calls the strongest open base model. In other words, Microsoft estimates its training method would need far less computing power to match that model. It says DeepSeek-V4-Pro-Base was chosen because it is the only released base model with more than one trillion parameters, and that Microsoft “put that saving into training a larger model.”
None of these results has been checked by outside researchers.
How it was trained
Microsoft stresses where its data comes from. “Our path to frontier performance is built on organic data, not distillation from other models,” the post says. Distillation means training a model on the outputs of another model.
According to the post:
- More data. The training corpus grew by “roughly 36%” in unique tokens compared with MAI-Base-1.
- Newer knowledge. The model has a March 2026 knowledge cutoff.
- Vision built in. Unlike its predecessor, MAI-Base-2 was pre-trained natively with images, using an “encoder-free” design that handles images inside the main model rather than through a separate component.
- Image data at scale. The image-and-text corpus holds 4.5 trillion unique tokens and made up 18% of a 50-trillion-token training mix.
Microsoft also credits “scaling ladders,” smaller experiments used to predict how a full-size model will perform, for training a 2.8-trillion-parameter model reliably.
Licensing through Foundry
Microsoft says its control over training data “will enable us to license our model weights to trusted partners through Foundry,” its enterprise AI platform. It says the models will be tuned first for Microsoft’s own products and then for Foundry customers. It has not said which partners, on what terms or when.
MAI-Base-3 is next
“While MAI-Base-2 goes through post-training, preparations for MAI-Base-3 are already underway,” the post says. Microsoft gave no size or timeline for the next model.
What’s still unknown
Microsoft has not said when a product built on MAI-Base-2 will launch, how it will be priced, or how it compares with leading proprietary models. Independent benchmarks will only be possible once the model is available.
What to watch
Watch for the post-trained models built on MAI-Base-2, any Foundry licensing deals and news on MAI-Base-3.
What this means
- Microsoft scales up its own models. MAI-Base-2 has 2.8 trillion parameters, with 80 billion active.
- Not yet usable. It is still in post-training, with no release date.
- Claims to check. All benchmarks so far are Microsoft’s own.
Sources
Related stories
- The 13 Billion Bet: Why Microsoft Is Building Its Own AI Stack
- Microsoft Unveils Maia 300 and Plans Multi-Gigawatt Maia Deployments From 2027; Maia 200 Is Now in Production
- Microsoft’s Sovereign AI Pitch: Not One Model, but the Right Model for Each Job
Follow @tsnmediahub on X for breaking AI news as it happens.
Follow on X
