Microsoft AI completes MAI-Base-2, a model with 2,800 billion parameters trained without distillation

Microsoft’s new foundation model, an MoE with 80 billion active parameters, is said to be 4 to 12 times more efficient during training than DeepSeek-V4-Pro-Base. Its weights will be made available via Foundry.

Microsoft AI has completed the pre-training of MAI-Base-2, an expert-mix model that has 2,800 billion parameters in total, 80 billion of which are active. It succeeds MAI-Base-1, on which MAI-Thinking-1 was built in June. The training mobilized 16,000 GB200 GPUs for 38 days, in August and September 2026, with an average utilization rate of 21% and effective training progress 89% of the time.

More efficient than the open source reference

On evaluations that are kept separate from training data (internal code, science, mathematics, Q&A ordered from vendors), MAI-Base-2 scores better than all the open models tested. Microsoft estimates that its training recipe requires 4 to 12 times less computation than that of DeepSeek-V4-Pro-Base to reach the same level, and that it has reinvested this gain in a larger model. The team attributes its ability to scale up unsurprisingly to a "scale" of models of increasing sizes and a technique that keeps the number of data repeats constant at each rung.

Microsoft insists on the origin of its data: pre-workout is based on "organic" data, without distillation of other models. The company intends to use this as an argument to license weights to trusted partners via its Foundry platform. The model is currently in post-workout.