Microsoft finishes pre-training MAI-Base-2, a 2.8-trillion-parameter model built fully in-house
First seen on X 7 hours ago@mustafasuleyman ♥ 1,961
Microsoft AI has finished pre-training MAI-Base-2, a very large base model built entirely in-house.
Microsoft AI CEO Mustafa Suleyman wrote on X on Oct 15 that the team now has "state-of-the-art pre-trained base models in the 3T-parameter weight class," with post-training results coming soon. Microsoft AI's technical blog says MAI-Base-2 is a mixture-of-experts model with 2.8 trillion total and 80 billion active parameters. Microsoft says it beats every open-weight model it tested on held-out evaluations and is 4-12x more pre-training-efficient than DeepSeek-V4-Pro-Base, the strongest open base model.
Suleyman added that on held-out code it matched DeepSeek-V4-Pro-Base's score with 12x less training compute. In June Microsoft released MAI-Thinking-1, post-trained on the earlier MAI-Base-1.
“We now have state-of-the-art pre-trained base models in the 3T-parameter weight class, built entirely in-house at MAI. Post-training results coming soon -- stay tuned!”