MAI-Base-2 gains efficiency and targets Microsoft’s next generation of models
Microsoft says it has completed pre-training of MAI-Base-2, a foundational model with 2.8 trillion parameters. The company highlights better computational efficiency, a multimodal base, and future distribution via Foundry.
Microsoft announced on October 15 that it had completed the pre-training of MAI-Base-2, its new foundation model, after 38 days on 16,000 Nvidia GB200 GPUs. The group claims to have significantly reduced the computational cost compared to several competing models, while preparing a base for its upcoming versions. Behind the announcement, the challenge is simple: to produce larger models, but above all more efficient ones. It’s not just a race for GPUs, even if the electricity bill is no small matter.
A larger model, but above all more computationally efficient
MAI-Base-2 is based on a mixture of experts architecture, or MoE for Mixture of Experts: only certain sub-models are activated for each request. Microsoft states that the entire model has 2.8 trillion parameters, of which 80 billion are active. According to its measurements, the model requires between 4.2 and 12.3 times fewer floating-point operations than DeepSeek-V4-Pro-Base to achieve comparable performance levels on several evaluations.
The publisher adds that computational efficiency improves by a factor of 6.3 to 12.7 compared to MAI-Base-1. In other words, Microsoft did not just stack parameters; it sought to better utilize each accelerator, which matters as much as raw size in a distributed infrastructure of this scale.
Why Microsoft is betting so much on architecture
To improve performance, the company says it worked on architecture, training algorithms, and hardware infrastructure. The distribution of experts was optimized to limit load imbalances, and the effective utilization of MoE layer computing capacity increased from 24% to 35% according to the figures provided. The model was able to apply this architecture at every layer.
Microsoft also adopted Muon, an optimizer intended to stabilize training and outperform AdamW in certain contexts. After optimization, Muon accounts for 2.1% of the average time per training iteration. During MAI-Base-2 training, the average utilization of the theoretical computing capacity of GPUs reached 21%, while training effectively progressed 89% of the time. In other words: lots of hardware, lots of software, and still quite some room for improvement.
Larger, cleaner, and more visual data
The quality of the corpus was also reworked. Microsoft says it increased the volume of unique training data by about 36% compared to MAI-Base-1, with more deduplication, classification, and content selection. The total dataset reaches 50 trillion tokens, of which 18% are related to vision-language data.
The model was natively pre-trained with image processing capabilities, according to a vision-language model architecture, or VLM. The publisher cites uses for understanding graphics, spatial reasoning, document analysis, medical computing, and interface interaction. Microsoft also states that the corpus is based on data collected and processed internally, without distillation from other models, and relies on an indexing robot respecting the robots.txt file.
What Microsoft is preparing now
The goal at this stage is as much industrial as technical: to control the entire chain, from data to infrastructure, to adjust models and execution systems simultaneously. Microsoft plans to offer the weights of its models to trusted partners via Microsoft Foundry, to facilitate their integration into enterprise applications and IT environments.
The exact terms remain to be specified: schedule, license, and resources needed for MAI-Base-2 deployment are not detailed. However, the publisher indicates that the model is already in post-training, a step that adapts it to concrete tasks, and that MAI-Base-3 is already in preparation. Time will tell if the promise of efficiency holds as well in real use as in benchmark tables.
Key points
- MAI-Base-2 was pre-trained in 38 days.
- Microsoft mobilized 16,000 Nvidia GB200 GPUs.
- The model has 2.8 trillion parameters.
- The vision-language share reaches 18% of the corpus.
- Microsoft is already preparing MAI-Base-3.
- Foundry is to distribute the weights to partners.
By the numbers
- 38 days — duration of pre-training, according to Microsoft, in August-September 2026.
- 16,000 Nvidia GB200 GPUs — infrastructure mobilized for MAI-Base-2.
- 2.8 trillion — total number of model parameters.
- 50 trillion tokens — size of the training mix, including 18% vision-language.
- 36% — increase in unique data volume compared to MAI-Base-1.