Microsoft AI on October 15 published the pretraining details of MAI-Base-2, its next-generation foundation model. It is a mixture-of-experts (MoE) model with 2.8 trillion total parameters and 80 billion active. Pretraining is finished and post-training is under way. MAI-Thinking-1, released in June, was built on the previous generation, MAI-Base-1.
Microsoft makes two main claims. The first is that "MAI-Base-2 outperforms every open-weight model we tested on our held-out evaluations." The second is efficiency: compared with DeepSeek-V4-Pro-Base, which Microsoft regards as the strongest open-source base model, MAI-Base-2 needs 4.2 to 12.3 times less pretraining compute to reach the same level, and 6.3 to 12.7 times less than Microsoft's own MAI-Base-1. The compute saved went into training a larger model. The 2.4T-A95B version of Qwen3.8 was also part of the comparison.
The evaluations use Microsoft's own held-out sets. The questions and item-level scores have not been released, so outsiders cannot reproduce the results for now.
Data and training
Microsoft states in its blog that all training data was processed in-house by the team and "does not involve distillation from other models." The training mix totals 50 trillion tokens, and by unique tokens the corpus is about 36% larger than that of MAI-Base-1. Visual data was added natively: 4.5 trillion unique visual tokens, making up 18% of the training mix.
The hardware was 16,000 GB200 GPUs, running continuously for 38 days between August and September 2026. Average MFU (model FLOPs utilization) was 21%, and effective training time was 89% of the total. The engineering changes centered on the MoE layers and the optimizer: MoE-layer MFU rose from 24% to 35%; the Muon optimizer was sharded and its communication optimized, cutting per-step overhead to 2.1% of iteration time, roughly 165 milliseconds; and training stability relied on activation gradient clipping, which Microsoft says doubled as a diagnostic tool for finding problems in the training data.
On method, the team used a scaling ladder to evaluate model changes at small scale and proposed source scaling to make sure conclusions from small experiments carry over to large scale. It also re-derived the scaling laws for learning rate, weight decay and dropout.
A back-of-envelope check
The numbers in the blog are enough to run a check yourself. Taking the BF16 theoretical peak of a GB200 as 2.5 PFLOPS, as the blog states, 16,000 GPUs at 21% utilization deliver about 8.4×10^18 floating-point operations per second of effective compute. Thirty-eight days at 89% effective time is about 2.92 million seconds. Multiplying the two gives a rough total training compute of about 2.5×10^25 FLOPs.
A common rule of thumb offers a cross-check: training compute is roughly 6 × active parameters × training tokens, or 6 × 80 billion × 50 trillion, about 2.4×10^25. The two methods land in the same order of magnitude, within 5% of each other, which suggests the MFU, duration and data volume Microsoft published are consistent with one another. In GPU hours, that comes to about 14.6 million.
Among models that have published their training figures, this scale ranks near the top. One piece of the comparison is still missing: Microsoft does not say how it estimated the compute of DeepSeek-V4-Pro-Base, and the blog does not explain how the baseline behind the "4 to 12 times" figure was chosen, so for now Microsoft's own framing is all there is to go on.
What is still unsaid
Post-training results for MAI-Base-2 will be announced once it is finished. Microsoft has not said whether it will release the weights, or in what product form the model will ship, whether built into Copilot or offered to enterprises through Foundry. None of that is visible yet.
The blog closes by noting that MAI-Base-3 is already in preparation. From MAI-Base-1 to 2, Microsoft grew the corpus by more than a third and pushed parameters to 2.8 trillion, while also publishing figures such as MFU and GPU hours that most labs prefer not to disclose. That level of transparency is rare among frontier labs.
Sources: Microsoft AI official blog, CocoLoop; parameter scale, training tokens, GPU count and MFU basis were checked, while total training compute is this site's rough estimate from the published figures.