Microsoft says MAI-Base-2 pre-training is complete; the 2.8-trillion-parameter model moves to post-training

Microsoft/MAI Superintelligence Models

Microsoft says pre-training of MAI-Base-2 is complete and the model is now in post-training, with results promised soon. It is a mixture-of-experts design: only a fraction of its weights run for each token. Microsoft also says preparations for a successor, MAI-Base-3, have started, without a date.

• Size: 80 billion active parameters out of 2.8 trillion total.
• Training ran on 16k GB200 GPUs over 38 days in August and September 2026, with average model FLOPs utilisation of 21% and goodput of 89%.
• No release date for MAI-Base-2 is given; Microsoft plans to license its weights to trusted partners through Foundry, timing not stated.
• MAI-Base-3 preparations are underway; no date given.