Microsoft completes pre-training of 2.8-trillion-parameter MAI-Base-2

Microsoft says it has completed pre-training MAI-Base-2, a mixture-of-experts model with 2.8 trillion parameters in total and 80 billion active. It is now in post-training, with enterprise licensing planned through Microsoft Foundry.

Microsoft AI Watch analysis

What happened

The model combines a shared expert with sliding-window hybrid attention and native multimodality. Microsoft says it was trained on roughly 50 trillion tokens over 38 days using 16,000 Nvidia GB200 GPUs. The company describes the training data as organic.

Microsoft also claims MAI-Base-2 is four to 12 times more FLOPs-efficient than DeepSeek-V4-Pro-Base. That comparison is Microsoft’s claim, not an independently established result. The company has not yet finished post-training, so the announcement is about a completed pre-training stage, not a model ready for customers.

Microsoft’s announcement says preparations for a MAI-Base-3 training cycle are already under way.

Why it matters

This is a substantial model-training effort, and Microsoft’s plan to license the model through Foundry could give enterprise customers another option alongside models from external providers. The scale and claimed efficiency are notable, but useful comparisons will depend on published methods and tests on real workloads.

Our read

The headline numbers make MAI-Base-2 worth watching, not yet worth declaring a winner. The consequential next step is whether post-training produces a capable model that enterprise customers can actually access, and whether Microsoft’s efficiency claim holds up under scrutiny. A trillion parameters is impressive; a clear account of what the model can do is more useful.

What to watch

• When post-training finishes and the model becomes available through Foundry.
• What evaluation results and methods Microsoft publishes.
• Whether the claimed efficiency advantage appears in comparable, practical tests. Activity teaser Microsoft says it has finished pre-training MAI-Base-2, a mixture-of-experts model with 2.8 trillion parameters in total and 80 billion active. The company plans to license it to enterprise partners through Foundry, but post-training is still under way. Microsoft also claims the model is four to 12 times more FLOPs-efficient than DeepSeek-V4-Pro-Base. That is a sizeable claim; the useful test will be comparable results and what the finished model can do. Our Forum story looks at what is known and what remains to be shown.

Discussion spark: What evidence should Microsoft publish before enterprises treat MAI-Base-2’s claimed efficiency advantage as a reason to choose it?

not affiliated with or endorsed by Microsoft