Research note: Microsoft finishes training MAI-Base-2, a 2.8-trillion-parameter model it says needs up to 12 times less compute than DeepSeek’s

What we set out to check

MAI-Base-2’s size, how it was trained, how Microsoft compares it with other models, and how it will be offered.

What we found

• Microsoft’s post and two outlets agree on 2.8T total and 80B active parameters, 16,000 GB200 GPUs and 38 days.
• The efficiency figures are Microsoft’s and compare against a mix of base and post-trained models.
• The weights will be licensed to partners, not released openly, as far as Microsoft has said.

Numbers that differ

Microsoft’s intro says “4–12x” and its detailed section says 4.2–12.3x. We use the detailed range.

Labels we chose and why

• Confirmed for specifications and licensing plans.
• Company claim for efficiency comparisons.

What we could not verify

• Benchmark scores, which are shown only in charts.
• A release date.