Microsoft builds a giant 2.8 trillion parameter AI

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 stage of training a very large AI model called MAI-Base-2. The model has 2.8 trillion parameters. That shows Microsoft is building its own top-tier AI, not only relying on partners.

What happened

On October 15, Microsoft AI published a technical post about MAI-Base-2. It is a "mixture of experts" model. That means it is built from many smaller specialist parts, and only some of them switch on for each piece of text. Microsoft says the model has 2.8 trillion parameters in total, and 80 billion are active at a time. Parameters are the adjustable numbers inside an AI that hold what it has learned.

The post says Microsoft recently finished "pre-training," which is the long first stage where the model reads huge amounts of text. The model is now in post-training, the polishing stage. Microsoft did not share post-training results yet, so nobody outside the company can test it.

Microsoft also made an efficiency claim. It says MAI-Base-2 is 4 to 12 times more efficient to pre-train than DeepSeek-V4-Pro-Base, which it calls the strongest open-weight base model. The details matter, so read the next section.

Key facts

• Training ran on 16,000 GB200 chips for 38 days in August and September 2026, according to Microsoft AI.
• The training mix was 50 trillion tokens. A token is a small chunk of text.
• Microsoft says all training data is organic, not made by other AI models.
• The efficiency range in the intro is 4 to 12 times. A later section gives 4.2 to 12.3 times, and we could not find an explanation for the small gap in the post.
• Microsoft says these numbers are estimates from scaling laws fitted on four held-out tests.
• Microsoft says it plans to license the weights to trusted partners through Foundry. It has not said it will release them to the public.

Background

The post says MAI-Base-2 follows MAI-Base-1. Microsoft says MAI-Thinking-1, announced in June 2026, was built on top of MAI-Base-1. It also says work on MAI-Base-3 is already underway.

The post compares MAI-Base-2 to other big open-weight models: DeepSeek-V4-Pro-Base, Kimi K3, Qwen3.8-2.4T-A95B, and MiMo-V2.6-Pro-RL. Microsoft says its model scored lower (better) on a test called bits-per-byte than all of them. It adds that the last three were tested only in their finished, post-trained form, which may change the comparison.

Techmeme linked the post in a short item with the size, chip and day numbers.

What it means for you

You cannot use MAI-Base-2 today. A base model is a starting point, not a chatbot. It still has to be tuned before it can answer questions.

Our take: the bigger shift is about who builds AI. Microsoft has leaned on a partner for much of its AI. A model of this size, trained on its own, suggests it wants a strong option of its own. If it works, you may see it in Microsoft products over time.

Our take: treat the efficiency claim as Microsoft's own estimate until others test it.

What happens next

Per Microsoft AI, post-training is underway and the weights are planned for trusted partners through Foundry. Microsoft gave no date. Our take: watch for post-training results, which will show how good the model really is.

FAQ

What does 2.8 trillion parameters mean?

Parameters are the numbers inside an AI that store what it learned. Microsoft says MAI-Base-2 has 2.8 trillion, but only 80 billion are active at one time.

Is MAI-Base-2 finished?

Not yet. Microsoft says pre-training is done and post-training is underway.

Can I use it?

Not now. Microsoft only says it plans to license the weights to trusted partners through Foundry.

Sources

• Microsoft AI
• techmeme.com
• techmeme.com