Microsoft finishes pre-training MAI-Base-2, a 2.8-trillion-parameter model

Microsoft AI, the company's in-house AI lab, said on 15 October 2026 that it has finished pre-training MAI-Base-2, a new foundation model with 2.8 trillion parameters. The model is now in post-training, the stage that turns a raw base model into a usable assistant. Microsoft has not given a release date or price.

In a long technical post, Microsoft explains how it trained the model and says it plans to license the weights to "trusted partners" through its Foundry platform, rather than publish them openly.

The model in numbers

According to Microsoft's post:

• Size: MAI-Base-2 is a mixture-of-experts (MoE) model with 2.8 trillion total parameters, of which 80 billion are active for each token. An MoE model only uses a small part of its network for each piece of text, which saves computing power.
• Training run: it was trained on 16,000 NVIDIA GB200 GPUs over 38 days in August and September 2026. Microsoft reports that training was making progress 89% of the time.
• Data: the training mix totalled 50 trillion tokens, with web data up to a March 2026 knowledge cutoff. Microsoft says its pool of unique training tokens grew by about 36% compared with MAI-Base-1.
• Vision from the start: unlike MAI-Base-1, which only read text, MAI-Base-2 learned from images during pre-training. Image-and-text data made up 18% of the training mix and covers tasks such as chart reading, medical images, computer use and even reading sheet music.

For comparison, Technology.org notes that MAI-Base-1, the base for the MAI-Thinking-1 reasoning model announced in June, had 1 trillion total parameters.

Microsoft's performance claims

Microsoft says MAI-Base-2 "outperforms every open-weight model we tested" on held-out tests. It measured how well each model predicts unseen text in four areas: code, STEM, maths, and questions and answers. For code it used its own internal codebase, and it commissioned the other test sets from outside vendors to reduce the chance the models had already seen them.

The company also says its training recipe is 4.2 to 12.3 times more compute-efficient than the one behind DeepSeek-V4-Pro-Base, which it calls the strongest open base model.

There are important limits to these claims, and Microsoft states some of them itself:

• Only open models were compared. No closed models such as OpenAI's or Google's appear in the published results.
• Base models were compared with finished models. DeepSeek is the only rival that released a base checkpoint. Kimi K3, Qwen3.8 and MiMo were tested as post-trained models, which Microsoft says may affect the results.
• The efficiency figure is an estimate. It compares training recipes using fitted scaling curves, not two finished models side by side.

These are Microsoft's own measurements and have not been independently checked.

How Microsoft built it

The post describes several technical changes. Microsoft says it fixed inefficiencies that allowed it to use expert layers in every block of the network, switched to the Muon optimizer, and added a method it calls activation gradient clipping to keep training stable. It also says the technique helped it find a bug in its data pipeline.

Microsoft stresses that the model was trained on "organic data" and does not use distillation, meaning it was not trained on the outputs of other companies' models. Its web crawler respects robots.txt, the file that tells crawlers which pages they may use, according to the company.

Why it matters

MAI-Base-2 is the largest step so far in Microsoft's effort to build its own frontier models alongside its partnership with OpenAI. The plan to license the weights through Foundry points to a business model aimed at companies that want to run a large model under their own control, with a clear record of where its training data came from.

For now, though, nobody outside Microsoft can use MAI-Base-2. The real test will come when the post-trained version is released. Microsoft says work on MAI-Base-3 has already started.

Sources

• Microsoft AI: Frontier Pre-training at Microsoft AI
• Microsoft AI: Frontier Pre-training at Microsoft AI (PDF)
• Technology.org: Microsoft finishes pre-training MAI-Base-2, a 2.8-trillion-parameter model
• StreetInsider: Microsoft completes pre-training of MAI-Base-2 AI foundation model