Microsoft finishes training its giant MAI-Base-2

Corrected 19 Oct: we fixed details of this story after checking the source again.

Microsoft's AI team says it has finished the first big training stage of a new model called MAI-Base-2. The company says it is very large and cheaper to train than a leading rival. If that holds up, it shows Microsoft can build its own top-level AI instead of relying only on partners.

What happened

On October 15, 2026, Microsoft AI published a post about MAI-Base-2. It calls the model a foundation model. That means a base model that other, more finished models get built from.

Microsoft says the model has 2.8 trillion parameters in total. Parameters are the settings inside an AI model, and more of them usually means a bigger model. The model uses a design called Mixture-of-Experts. Only part of the model switches on for each task. Microsoft says 80 billion parameters are active at a time.

The company says it trained the model on 16,000 GB200 chips over 38 days, in August and September 2026. GB200 is a type of Nvidia AI chip. Microsoft says the model beats every open-weight model it tested on its own held-back tests. Open-weight means anyone can download the finished model files. These tests are Microsoft's own, and the post is the only source for these results.

Key facts

• Size: 2.8 trillion parameters in total, with 80 billion active at a time, per Microsoft AI.
• Training time: 38 days in August and September 2026, on 16,000 GB200 chips, per Microsoft AI.
• Efficiency claim: Microsoft says the model is 4 to 12 times more efficient to train than DeepSeek-V4-Pro-Base, which it calls the strongest open base model. Its longer analysis gives a range of 4.2 to 12.3 times.
• Data: Microsoft says it used real, collected data and did not copy answers from other AI models. It says the data was last updated in March 2026.
• Status: the model is now in post-training, the stage that turns a base model into a finished one.

Background

Microsoft AI has been building its own models for a while. The post says an earlier model, MAI-Thinking-1, was announced in June 2026 and was built on an earlier base model called MAI-Base-1. Microsoft says MAI-Base-2 is 6.3 to 12.7 times more efficient than that first base model.

Microsoft also says its training data has a clear history, which it calls strong data provenance. It presents that as a reason to trust the model. The post says its training data has about 36 percent more unique tokens than MAI-Base-1 had. Tokens are small chunks of text.

What it means for you

Most people will not use a base model directly. It is the raw engine. But this one will likely show up later as part of Microsoft products, because the company says it will first use the model in its own products.

Our take: a claim about being 4 to 12 times more efficient is a company claim. Wait for outside testers to repeat the results before you treat it as proven. Efficiency matters because cheaper training can mean cheaper tools for everyday users over time.

Microsoft also says it plans to license the model weights to trusted partners through its Foundry service. So businesses may be able to use it, even if you never see it yourself.

What happens next

Microsoft says post-training is under way and that it will share more results soon. It also says preparations for a third model, MAI-Base-3, have started. No dates are given in the post.

Our take: watch for the finished version and for independent test scores. Those will show whether the efficiency claim is real.

FAQ

What is MAI-Base-2?

It is a large base AI model from Microsoft AI. Microsoft says it has 2.8 trillion parameters in total and finished its first training stage after 38 days, per its own post.

Can I use MAI-Base-2 now?

Not yet. Microsoft says the model is still in post-training and that it plans to use it in its own products first. It also plans to license the weights to trusted partners.

Is it better than DeepSeek?

Microsoft says it is 4 to 12 times more efficient to train than DeepSeek-V4-Pro-Base, and that it beat every open-weight model it tested. Those are Microsoft's own tests, so outside checks are still needed.

Sources

• Microsoft AI