Microsoft trains 2.8 trillion-parameter AI model with up to 12 times less compute
Stockwhiz•15/10/2026•12:34 ET
20
Key Highlights
- ➤MAI-Base-2 totals 2.8 trillion parameters, with 80 billion active per task
- ➤Microsoft trained MAI-Base-2 on 16,000 Nvidia (NASDAQ: NVDA) GB200 GPUs over 38 days
- ➤MAI-Base-2 was 4.2 to 12.3 times more compute-efficient than DeepSeek-V4-Pro-Base
- ➤MAI-Base-2 was 6.3 to 12.7 times more compute-efficient than Microsoft’s MAI-Base-1
- ➤MAI-Base-2 dataset grew about 36%, adding native vision training
Microsoft (NASDAQ: MSFT) completed pre-training of MAI-Base-2, a 2.8 trillion-parameter model it said beat every open-weight model tested and required up to 12 times less computing power to match a leading rival.
MAI-Base-2 has 80 billion active parameters and uses a mixture-of-experts architecture, which activates only a portion of its total parameters for each task. The model is now undergoing post-training, with Microsoft saying further results are expected later.
Get started
Create a free account to read the full story.