🔍 Intelligence Synthesis JP/EN
Ethics · Regulation · Technology — Pharma Practice Notes

AI Tech Flash ── 2026-10-17

Microsoft AI finishes pre-training MAI-Base-2

Microsoft AI unveils a 2.8-trillion-parameter MoE foundation model

What changes

According to Microsoft AI's announcement, the company now has a large foundation model it trained itself. A foundation model is the base AI model, before it is tuned for uses such as chat.

The company says it used a method that needs less computing power for training. If confirmed, building large models could cost less. No independent party has verified this yet.

Why it matters

According to Business Insider, Microsoft has long used OpenAI's models in its own products. The relationship between the two companies has changed over the past year or so, and Microsoft has been working on its own models.

Tech Insider called the announcement the clearest signal yet that Microsoft wants its own model. The company also claims to beat DeepSeek on training compute efficiency. It has not published the benchmark scores themselves.

What it is

MAI-Base-2 is a foundation model whose pre-training Microsoft AI has finished. Pre-training is the first stage of building a model, when it learns patterns from large amounts of text and images.

According to the company, it has 2.8 trillion total parameters, with 80 billion active per token. Parameters are the numbers adjusted during training. A token is a small unit of text or image data the model processes.

The model is a mixture-of-experts (MoE) design. Only part of the network runs for each input.

Who and where

Microsoft AI built it. According to Tech Insider, pre-training ran on 16,000 Nvidia GB200 GPUs for 38 days, in August and September 2026.

The model is still in post-training, the stage after pre-training in which it is tuned for specific uses. No release date or pricing has been announced, and no chatbot has been announced. Tech Insider's summary table lists a plan to license it to partners through Azure AI Foundry.

Why it was built

According to Business Insider, pre-training takes a great deal of costly computing power. A more efficient method could save Microsoft money.

Microsoft AI claims its recipe is more efficient than DeepSeek-V4-Pro-Base. It says it put those savings into training a larger model.

How it works

In an MoE model, only 80 billion of the 2.8 trillion parameters are active for each token. Tech Insider says this keeps the cost of inference down. Inference means running a trained model.

According to Microsoft AI, the model was trained with vision built in from the start, using an encoder-free vision-language design. Tech Insider reports about 50 trillion training tokens, roughly 18% of them visual. That figure comes from secondary reporting.

Key points

  1. According to Microsoft AI, it is an MoE model with 2.8 trillion total parameters and 80 billion active per token.
  2. The company claims its pre-training was 4.2 to 12.3 times more compute-efficient than DeepSeek-V4-Pro-Base. That range covers four evaluations.
  3. Post-training is still under way. No release date or pricing has been announced, and no benchmark scores have been published.

Conclusion

MAI-Base-2 is a large foundation model whose pre-training Microsoft AI has finished. The efficiency and performance claims so far come from the company.

Mustafa Suleyman's comparison with OpenAI was made at an internal town hall. Until outside parties can test the model, the figures cannot be judged.

Quotes and sources

“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,”
── source, 2026-10-16
“It is 4-12x more pre-training FLOPs-efficient than the strongest open base model, DeepSeek-V4-Pro-Base, and we have put that saving into training a larger model.”
── source, 2026-10-16
“We now have a stronger pre-trained base model than the model that currently powers GPT 5.6 Sol,”
── Business Insider, 2026-10-15
“Pre-training is the first stage of building an AI model, when it learns patterns from huge amounts of text, images, and other information.”
── Business Insider, 2026-10-15

Source articles

Written by AI (Claude) from the cited articles, then checked by Jev (TypeSafe's judgement model), a Gemini fact-check and a verbatim check of every quote.