Microsoft is making its way back into the frontier model race.
On Thursday, the company's AI research lab announced it had completed pretraining on MAI-Base-2, a mixture-of-experts foundation model with 2.8 trillion parameters and 80 billion active. Additionally, the model outperforms a number of open-weight models on several evaluations, and is far more efficient than the most powerful open model it was tested against, DeepSeek-V4-Pro-Base.
The company tested the model on "held out" evaluations, meaning it was tested on a subset of data that was excluded from the training process, in subjects such as coding, math, STEM, and question answering.
- Though it didn't test against models from OpenAI, Anthropic, Google or other frontier labs, MAI-Base-2 beat out trillion-plus parameter models from Alibaba, DeepSeek, Moonshot, and Xiaomi in these metrics.
- As for efficiency, the company said its model is four to 12 times more "Pre-training FLOPs-efficient" than DeepSeek's model, meaning that it used far less raw computation for internal training to achieve similar performance, largely as a result of better training data and changes to model architecture. Microsoft said it directed this cost savings into a larger model.
"Owning our model pipeline end-to-end allows Microsoft to make a distinctive offering to our customers," Microsoft said in its research blog. "Microsoft has strong data provenance over training tokens that go into our model. Our pre-training is built on organic data which does not involve distillation from other models."
Though Microsoft's AI strategy has differed from competitors like OpenAI, Anthropic and Google, according to Business Insider, Microsoft AI Chief Mustafa Suleyman told employees in a town hall on Thursday that the company's unnamed in-house "pretrained base model" performed better than a OpenAI's GPT 5.6 Sol on private testing, and used "one-quarter of the compute" to achieve that result.
Our Deeper View #
Microsoft has taken a unique strategy in the AI race compared to its frontier lab competitors. Rather than diverting tons of resources and energy to fighting for more powerful AI and more attention against major model providers, it's taken a more conservative approach, focusing on safety, security, and efficiency over brute-force scaling. It's a prudent strategy that's evident in the essays from both Suleyman and Microsoft CEO Satya Nadella, calling for caution around things like model observability and AI consciousness. It's also evident in the company's prior model releases themselves, which are focused largely on cost and resource efficiency that pass savings down to the enterprises that use them. MAI-Base-2, while certainly closer to a frontier model in size than those it's released in recent months, is another example of the company putting efficiency first, and highlights that Microsoft has its finger on the pulse of the things that are most important to its enterprise customer base: Security and cost savings.