MicrosoftAnalysisMicrosoft AI details 2.8T-parameter MAI-Base-2 foundation model
What happened.Microsoft AI completed pre-training on MAI-Base-2, a 2.8-trillion-parameter mixture-of-experts model activating 80 billion parameters per token.
Why it matters.Hyperscalers are building proprietary foundation weights to reduce dependency on third-party model providers. Builders using Microsoft Foundry will gain direct access to custom enterprise model tuning.
The technical specifications:
- Compute cluster: Engineers trained the checkpoint across 16,000 NVIDIA GB200 GPUs over 38 days using Muon optimization to cut training latency.
- Multimodal corpus: Ingested 50 trillion tokens of organic non-distilled data, including 4.5 trillion unique tokens dedicated to native vision-language reasoning.
- Compute efficiency: Yields between 4 times and 12 times higher training FLOP efficiency than leading open trillion-parameter base models.