Microsoft says it completed pre-training of MAI-Base-2

Microsoft AI said on October 15 that it had completed pre-training MAI-Base-2 using 16k GB200 graphics chips over 38 days in August and September. The model activates 80 billion of its 2.8 trillion parameters for each task, using a design that switches on only part of the network. Microsoft plans to license the model weights to trusted partners through Foundry after further training. AI chief Mustafa Suleyman reportedly told staff it outperformed GPT-5.6 using about a quarter of the computing resources, but that comparison rests on private tests. Microsoft supplied no public benchmarks, pricing or release date.