Microsoft builds a 2.8T-parameter model of its own

Microsoft is training frontier-scale models of its own now, and it wants you to know the run was cheap for its size.

Microsoft AI said on Oct. 15 it has finished pre-training MAI-Base-2, a mixture-of-experts model with 2.8 trillion total parameters and 80 billion active. The model is now in post-training and has no release date.

What Microsoft says it built

The training mixture was 50 trillion tokens, including 4.5 trillion unique vision tokens, with a knowledge cutoff of March 2026. Microsoft says it processed every token in-house and did not distill from other models.

The run took 38 days in August and September on 16,000 GB200 GPUs, at 21% average model FLOPs utilization and 89% goodput, per the post.

Efficiency is the claim Microsoft leads with. It says MAI-Base-2 is 4 to 12 times more pre-training compute-efficient than DeepSeek-V4-Pro-Base, and 6.3 to 12.7 times more than its own MAI-Base-1.

4 to 12 times.

Microsoft built those figures from its own held-out evaluations, including code from its own codebase and STEM, math and Q&A sets it commissioned from vendors, plus scaling-law fits. Some comparison models were tested only as post-trained checkpoints, which the post says may affect the results.

The full stack is the point

Nobody chats with a base model. It is raw material. What Microsoft is announcing is that it can make the raw material itself.

The post says owning the full stack lets Microsoft tune its models and infrastructure together "for the best performance at the lowest cost." Earlier this week it made Maia its main AI inference chip. A model of its own is the other half of that plan.

The no-distillation line is pointed too. Copying rivals' outputs is a live fight: this week Anthropic said it cut Claude distillation by up to 99%.

Nothing to use yet

Microsoft says the model will reach its own products first and then Foundry customers. It plans to license the weights to trusted partners through Foundry. The post mentions no public open-weight release.

So for now this is a roadmap item. If you build on Azure, add it to your list of models to evaluate once it ships, and hold any switching plans until there are numbers you can check.

If you compare base model claims for a living, note the method. Bits-per-byte on held-out sets and scaling-law fits are reasonable tools, but they are Microsoft's sets and Microsoft's fits.

What to watch

Watch for the post-trained model and its scores on public benchmarks, not Microsoft's own sets. Microsoft also says preparation for MAI-Base-3 is underway.

Efficient pre-training is the promise. The finished model is the proof.