Microsoft MAI-Base-2: 2.8T parameters, no price, no release date

Microsoft AI has finished pre-training MAI-Base-2, a 2.8-trillion-parameter model, and it plans to license the weights to "trusted partners" through Foundry. That is the headline from Microsoft's October 15 lab notes. It is a big number, and a real signal that Microsoft wants its own frontier stack. It is not a product yet. There is no price, no release date and no API to call.

What actually happened

Per Microsoft AI's post:

• The model. MAI-Base-2 is a mixture-of-experts base model with 2.8 trillion total parameters and 80 billion active per token.
• The run. 16,000 Nvidia GB200 GPUs for 38 days in August and September 2026, on a 50-trillion-token training mix.
• The efficiency claim. Microsoft says the model is 4.2x to 12.3x more pre-training FLOPs-efficient than DeepSeek-V4-Pro-Base. It reports 21% average model FLOPs utilization and 89% goodput. These are Microsoft's own numbers, not an independent test.
• Vision built in. Vision-language data is 18% of the mix, so the base model is multimodal from the start.
• Data provenance. Microsoft says it used organic data with no distillation from other models.
• Licensing. Microsoft plans to license the weights to trusted partners through Foundry. It did not name partners or publish terms.
• Status. The model is now in post-training. Microsoft says more results are coming "soon" and that work on MAI-Base-3 has started.

Why MAI-Base-2 matters for your business

If you run on Azure or Microsoft 365 Copilot, your model supplier is quietly becoming Microsoft itself, not only OpenAI. We already saw Microsoft swap the models under Copilot. A base model at this scale makes that path longer.

The "no distillation" line also matters. A vendor that can show where its training data came from has a cleaner licensing story. That is the pitch to enterprises who want weights in their own tenant.

But a pre-training post is a press release about a future product. What we would do:

• Do not plan around it. Build on models you can call today, with a published price.
• Keep your model layer swappable. One adapter per provider, prompts and evals in your repo. When MAI-Base-2 ships, you test it in a day.
• Ask your Azure rep the real questions. Who qualifies as a "trusted partner"? What does a weight license cost? Where do the weights run?
• Wait for independent evals. Vendor-run benchmarks are a starting point, not a decision.

Sources: Microsoft AI.