What was built
Microsoft AI said on 15 October that it has completed pre-training on MAI-Base-2, a mixture-of-experts foundation model with 2.8 trillion total parameters and 80 billion active. The company published the details in a lab note on its own site rather than as a product launch, and the model is not available: it is in post-training, with no checkpoint, no API, no price and no release date.
The run itself is the substance. Microsoft says it trained the model on 16,000 GB200 GPUs over 38 days in August and September 2026, averaging 21% model FLOPs utilisation and 89% goodput — the share of wall-clock time during which training actually progressed. Microsoft notes that the MFU figure is normalised against the GB200’s theoretical BF16 peak of 2.5 petaflops and excludes rematerialisation and optimiser FLOPs, which is the kind of qualification that usually goes unstated.
The training mixture totalled 50 trillion tokens. Vision-language data made up 18% of it once repetition is counted, drawn from a 4.5-trillion-token unique VLM corpus, and unlike its predecessor the model was pre-trained natively with vision rather than having sight bolted on afterwards. The unique-token corpus grew by roughly 36% over MAI-Base-1, with a knowledge cutoff of March 2026.
The number to read carefully
The headline claim is that MAI-Base-2 is 4.2 to 12.3 times more pre-training FLOPs-efficient than DeepSeek-V4-Pro-Base, which Microsoft describes as the strongest open base model it tested. Microsoft’s own summary rounds that to 4–12x.
That is not a head-to-head benchmark result. Microsoft arrived at it by fitting scaling laws on four held-out evaluations and comparing the implied compute against DeepSeek’s training compute. It is an estimate of how much compute a rival would need to reach the same scores, derived from curves Microsoft fitted itself, against a figure for DeepSeek’s spend that Microsoft does not control. The range is wide for the same reason: 4.2 and 12.3 are both outputs of the same method on different evaluations.
The company reports a 6.3 to 12.7 times improvement over its own MAI-Base-1 by the same method, which is the more informative comparison, because both sides of it are Microsoft’s and measured the same way.
Three levers, and no distillation
Microsoft attributes the gains to scaling science, model-infrastructure co-design and more data. The first covers scaling ladders that account for repeated training data, and empirically fitted hyperparameter scaling laws. The second covers improved mixture-of-experts and Muon optimiser implementations, plus activation gradient clipping. The third is the expanded web pipeline and the vision corpus.
The company also states that it did not distil from other models. That matters commercially as much as technically: a base model trained without another lab’s outputs is cleaner to license, and Microsoft says it plans to license the weights to trusted partners through Foundry.
Why Microsoft is building this at all
Microsoft pays OpenAI for frontier models and ships them across its products. Building a 2.8-trillion-parameter base in-house, on its own silicon allocation, is an option on not needing to. The lab note says preparations for MAI-Base-3 are already under way, which is the clearer signal: this is a programme, not a demonstration.
What to watch
Post-training is where a base model becomes something you can use, and where the claims become checkable. Until Microsoft ships a post-trained model that outside evaluators can run, the efficiency figures remain the company’s own estimates on its own held-out tests. Watch for the Foundry licensing terms, for any independent evaluation against DeepSeek-V4-Pro-Base, and for whether Microsoft publishes the four held-out evaluations it fitted the curves on.
