Microsoft has completed the pre-training phase of NEVER-Base-2its new language model based on Mixture of Experts architecture. The numbers are those that require a pause for reflection: 2,800 billion total parametersof which only 80 billion active for each inference, thanks to the MoE logic that routes each request to a restricted subset of “experts” internal to the model.
MAI-Base-2 takes one approach encoder-free for viewing capabilitiesnatively integrating visual processing into the model without resorting to a separate module trained separately, as is the case in many competing multimodal architectures.
To achieve this result, Microsoft has deployed non-trivial computing power: 16,000 GB200 GPUs committed to 38 daysbetween August and September 2026. During training, the average rate of use of available FLOPs (MFU) stopped at 21%, while the “goodput”, i.e. the fraction of time in which the training actually progressed without interruptions or slowdowns, reached 89%.
Training data and comparison with DeepSeek
The training mix includes 50 trillion tokens overall. Of these, the 18% it belongs to vision-language data, for a single multimodal corpus equal to 9,000 billion tokens: a significant share, which confirms how much Microsoft is betting on native multimodality rather than on a subsequent addition.
On the efficiency front, Microsoft claims that its training recipe results from 4.2 to 12.3 times more efficient in terms of FLOP compared to DeepSeek-V4-Pro-Base, one of the most cited references among Chinese open-weight models. The company also claims that MAI-Base-2 outperforms every open-weight model tested in its internal evaluations, but for now these are numbers communicated by the company itself, without confirmation from independent benchmarks.
Microsoft has not indicated either a release date or a price: the model is currently being released post-trainingthe next step that normally serves to transform a basic model into a usable conversational or agentic product. The company clarified that the model weights will not be open to the public, but only licensed to trusted partners through Foundry, Microsoft’s platform dedicated to distributing AI models to businesses.
The strategy recalls what has already been done with MAI-Thinking-1, announced in June 2026 as a reasoning model built on top of MAI-Base-1. Microsoft therefore seems to consolidate a cycle in which the basic model is produced internally and then refined for specific purposes, with simplistic access to weights that keeps it distant from the more open logic adopted by other laboratories. The comparison with DeepSeek, however flattering on paper, will have to be verified when third-party benchmarks arrive.
