Microsoft says MAI-Base-2 is 4 to 12 times more compute-efficient than DeepSeek's base model
Microsoft says no other model's outputs went into training it, and it wants to license the weights to businesses.
On October 15, 2026, Microsoft AI said it had finished pre-training MAI-Base-2, a mixture-of-experts model with 2.8 trillion parameters in total, 80 billion of them active. The lab put the results in a post on its own site.
The number it leads with is a comparison with a Chinese open model. Microsoft says MAI-Base-2 is 4 to 12 times more efficient in pre-training compute than DeepSeek-V4-Pro-Base, which it calls the strongest open base model, and that it spent the saving on training a bigger model.
- Size
- 80 billion active and 2.8 trillion total parameters
- Training
- 16k GB200 GPUs over 38 days in August and September 2026
- Main comparison
- DeepSeek-V4-Pro-Base
- Status
- currently undergoing post-training
What Microsoft compared it with
Microsoft scores the models on held-out text in bits per byte, which doesn't depend on each model's tokenizer. The STEM, math and Q&A sets were commissioned from vendors. The code set is MAI's own codebase.
On those tests, the post says, MAI-Base-2 beat every open-weight model Microsoft tried. DeepSeek's is the only other base model in the lineup, so it's the cleanest match. The rest come with a warning from Microsoft itself:
Only post-trained checkpoints are available for these models. We report their results with the caveat that post-training may affect these evaluations.
Closed models are missing. The post doesn't set MAI-Base-2 against anything from OpenAI, Google or Anthropic, so I wouldn't read it as a frontier result just yet.
How the run went
Microsoft trained it on 16,000 GB200 GPUs over 38 days in August and September. Training was making progress 89% of that time, a figure the post calls goodput, and the chips ran at 21% of their theoretical peak on average.
Its first model in this line, MAI-Base-1, was smaller than the open and closed models at the frontier. So the plan after it was simple: "our path ahead for pre-training was clear: scale up." Much of the post is about the tooling for that, like testing changes on small models first (a ladder of sizes trained the same way) before trusting them at 2.8 trillion.
Who gets to use it
For now, only Microsoft. It says it'll tune the model for its own products first and then for customers of Foundry, its platform for selling models to businesses.
Our pre-training is built on organic data which does not involve distillation from other models.
That line matters for the business plan. Microsoft says the clean data history will let it license the model's weights to trusted partners through Foundry, though the post names no partner and no date.
MAI-Base-2 is in post-training now. The post says only that Microsoft looks forward to sharing more results soon.