Microsoft trains MAI-Base-2, its most efficient frontier foundation model yet
The new MAI-Base-2 model uses far fewer computing resources than rival open-weight models while delivering state-of-the-art results.
A faster way to train giant AI
Microsoft has finished pre-training a new frontier foundation model called MAI-Base-2. The model belongs to a family that also includes MAI-Base-1 and MAI-Thinking-1, the reasoning-focused version Microsoft announced in June. Like its predecessor, MAI-Base-2 is built on a Mixture of Experts (MoE) design, a technique that spreads a massive model across many smaller "expert" networks so only a subset activates for each task.
MAI-Base-2 has 2.8 trillion total parameters but uses just 80 billion active ones during inference. That sparse approach means it can handle complex reasoning without paying the full computational cost of its enormous size.
Efficiency that scales
On held-out evaluations, Microsoft says MAI-Base-2 outperforms every open-weight model in its class. More striking is its efficiency: it requires 4 to 12 times fewer pre-training FLOPs (floating-point operations, the standard unit for measuring AI training compute) than DeepSeek-V4-Pro-Base, the strongest open base model Microsoft tested.
Microsoft is treating that efficiency gain not as an end in itself but as a resource to reinvest. The company says it has put the saved compute into training a larger model, suggesting future iterations will build directly on these gains.
Where it fits in the frontier
Frontier AI refers to the most capable models at the cutting edge of research, often developed by well-funded labs with access to enormous compute clusters. Microsoft's decision to share the training approach, even if it keeps the model weights private, fits a growing pattern of companies using published research to guide the field.
MAI-Base-2 slots into Microsoft's broader MAI model family. Earlier this year the company introduced MAI-Base-1, and in June released MAI-Thinking-1 as a post-trained reasoning layer on top of that base. Together the series covers both general-purpose and reasoning-optimized paths, giving developers a choice depending on the task.
What comes next
No public release date or access details for MAI-Base-2 have been announced. Microsoft has not disclosed whether the model will eventually become open-weight or remain proprietary. What is clear is that the efficiency improvements are measurable and substantial enough to reshape how the company plans its next training run.
Key facts
- MAI-Base-2 is an MoE model with 2.8 trillion total and 80 billion active parameters.
- It outperforms every open-weight model Microsoft tested on held-out evaluations.
- It is 4-12 times more pre-training FLOPs-efficient than DeepSeek-V4-Pro-Base.
- Microsoft is reinvesting compute savings into training a larger follow-on model.
- The model is part of the MAI family alongside MAI-Base-1 and MAI-Thinking-1.
Sources
- Frontier Pre-training at Microsoft AI | Microsoft AI — microsoft.ai
- Generative AI — en.wikipedia.org
- Foundation model — en.wikipedia.org
- Claude (AI) — en.wikipedia.org
This story was written before each paragraph carried a source number: every fact in it comes from one of the sources above.
Related stories
New modelsMicrosoft Unveils Microsoft-Decision-1 , a Decision AI Model Built on Alibaba 's Qwen3.5Herald of AI
Microsoft Unveils Microsoft-Decision-1 , a Decision AI Model Built on Alibaba 's Qwen3.5
A new kind of AI that picks from a list instead of writing text could make agents faster and cheaper.