New to this? Read it in simple words
- Microsoft finished the first, longest stage of training its biggest AI model, MAI-Base-2.
- It has 2.8 trillion parameters, but only 80 billion work on each token.
- Microsoft says it needs 4.2 to 12.3 times less computing power than DeepSeek’s model.
- The model is not released yet, and no outside group has tested it.
- Pre-training
- The first, longest stage of making an AI model, when it learns patterns from huge amounts of data.
- Base model
- A model after pre-training only. It still needs more training before it can work as an assistant.
- Mixture of experts
- A model built from many smaller parts, of which only a few work on each piece of text.
Microsoft makes some AI models of its own, called MAI, next to its partnership with OpenAI. On October 15, its AI team said it had finished the first and longest stage of training its largest one.
What Microsoft built
MAI-Base-2 is a base model. That is the result of pre-training, the long first stage in which a model learns patterns from huge amounts of text, images and other data. The model is now in post-training, the stage that turns it into a useful assistant.
The model has 2.8 trillion parameters, Microsoft says. It is a mixture-of-experts model, so only 80 billion of them work on each token. Unlike MAI-Base-1, its smaller predecessor from June, it also learned from images from the start.
Microsoft trained it on 16,000 Nvidia GB200 chips for 38 days in August and September 2026. Its training data runs up to March 2026.
Microsoft’s own estimates of October 15, against DeepSeek-V4-Pro-Base, on its four private tests; not checked by an outside group.
What Microsoft claims
Microsoft says MAI-Base-2 did better than every open-weight model it tested, on four tests of code, maths, science and general questions. The tests are private. The code test uses Microsoft’s own code, and outside firms made the other three for Microsoft.
Its main comparison is DeepSeek-V4-Pro-Base, which Microsoft calls the strongest open base model. Microsoft estimates that its own training method needs 4.2 to 12.3 times less computing power to reach the same results. The gap is largest on code and smallest on general questions.
These are estimates from Microsoft’s own calculations, not a race between two finished models. Some rivals were only available in their finished, post-trained versions, and Microsoft notes that this may affect the comparison.
When people can use it
Microsoft has not released the model, and gives no date or price. It says it will use the model first in its own products, and then for customers of Foundry, its cloud service for AI.
Microsoft also plans to license the model’s weights to “trusted partners” through Foundry. So, unlike an open-weight model, it will not be free for anyone to download.
Microsoft says the model did not learn by copying other AI models, and that its web crawler respects robots.txt, the file in which websites set rules for crawlers. Work on MAI-Base-3 has already started, it says.
Sources1
The OpenAI comparison
Mustafa Suleyman, who leads Microsoft AI, went further at a staff meeting on October 15, Business Insider reports. He said Microsoft now has a stronger base model than the one behind OpenAI’s GPT-5.6 Sol, trained with a quarter of OpenAI’s computing power.
Suleyman said the comparison was made privately, using information Microsoft has about OpenAI’s models. It can’t be checked from outside. Microsoft declined to comment to Business Insider, and OpenAI did not respond.
No independent group has tested MAI-Base-2 yet.
Sources2
Sources
Every fact in this story comes from the sources below. Open them to check our work.
- 1
- 2Research · October 15, 2026Microsoft AI chief says the company beat OpenAI at its own game, using a quarter of the computing power Business Insider
- 3Research · October 16, 2026Microsoft announces completion of pre-training for its 2.8 trillion-parameter foundational model, ‘MAI-Base-2’ GIGAZINE
- 4
Every result comes from Microsoft’s own private tests, and the efficiency figures are its estimates. The comparison with OpenAI rests on one report of a staff meeting, and no outside group has tested the model.
Microsoft finished pre-training a 2.8-trillion-parameter model it says needs far less computing power, though no outside group has checked it. How we rate the mood