❯ Microsoft finishes pre-training MAI-Base-2, a 2.8-trillion-parameter model it says is 4x to 12x more compute-efficient than DeepSeek-V4-Pro-Base

A bigger, sparser baseMicrosoft AI announced that it has completed pre-training MAI-Base-2, a mixture-of-experts model with 2.8 trillion total parameters and 80 billion active. Microsoft says it beats every open-weight model it tested on its held-out evaluations and is 4x to 12x more pre-training FLOPs-efficient than the strongest open base model, DeepSeek-V4-Pro-Base, with the savings put into training a larger model.

38 days on 16,000 GB200sMicrosoft says MAI-Base-2 was trained on 16,000 GB200 GPUs over 38 days in August and September, with training progressing 89% of the time. In June it released MAI-Thinking-1, post-trained on the previous base, MAI-Base-1. The new model is still in post-training and not yet available to use.

Clean data as the pitchMicrosoft stresses that the model’s pre-training data is organic and involves no distillation from other models, giving it clear provenance. It plans to license the weights to trusted partners through Foundry. Microsoft wants to make traceable training data a reason to choose it over open models, especially for enterprise customers who worry about copyright and compliance risk.

▮ SIGNALCompanies building their own large models now pitch “we didn’t distill anyone” as a feature, adding a “where did the data come from” test to enterprise procurement that open models do not always pass.