Microsoft AI finishes pre-training MAI-Base-2, a 2.8T-parameter base model, and puts it ahead of all open-weight models in its tests

發生什麼事

At the 3-trillion-parameter scale, MAI already has pre-trained base models of its own, developed without outside help, according to Microsoft AI CEO @mustafasuleyman; post-training results will be out soon. The official technical blog gives the training details: 38 days on 16,000 GB200 GPUs between August and September, 50 trillion tokens of data, a knowledge cutoff of March 2026, and vision data built in from the start.

Every open-weight model Microsoft tested finished behind MAI-Base-2 on the company's held-out evaluations, by its own account. Using DeepSeek-V4-Pro-Base, which Microsoft regards as the best open base model, as the yardstick, it estimates that its own training method would reach the same level with only 1/12.3 to 1/4.2 of that model's compute. Licensing the weights is possible because, according to Microsoft, pre-training used only natural data, with no distillation from other models (official technical blog).