Microsoft has completed pre-training MAI-Base-2 and published details of how it built the model. The Microsoft AI announcement describes a mixture-of-experts system with 2.8 trillion total parameters and 80 billion active parameters.
A base-model milestone with work still underway
MAI-Base-2 is currently undergoing post-training. The announcement therefore describes a training milestone and research results, rather than a broadly available finished assistant.
Microsoft says owning the model pipeline will let it adapt the models and infrastructure together, first for its own products and then for Foundry customers. It also outlines an intention to license weights to trusted partners through Foundry. That statement should not be read as an unrestricted public download or a release date for every customer.
Read the efficiency claim in context
Microsoft reports training on 16,000 GB200 GPUs over 38 days. It describes changes to scaling experiments, mixture-of-experts implementation, optimization and data processing, including the integration of vision data.
The company compares held-out text performance using bits per byte, a metric intended to avoid differences between tokenizers. It estimates that its training recipe needs roughly 4.2 to 12.3 times fewer FLOPs to match DeepSeek-V4-Pro-Base on four held-out evaluations.
This is Microsoft's estimate of training-recipe efficiency under its scaling analysis. It is not a claim that an application using MAI-Base-2 will answer every request that many times faster or at that fraction of the serving cost.
Check access and the next evaluation stage
The report distinguishes base-model evaluations from results affected by post-training. For several comparison models, Microsoft notes that only post-trained checkpoints were available.
For developers, the next useful evidence is the post-trained model's supported tasks, deployment conditions and application-level behavior. The pre-training report supplies technical context for those later decisions, but does not establish them on its own.