Microsoft AI, the company's AI lab, said on Thursday, Oct. 15, that a new model, MAI-Base-2, has finished pretraining, the first, broad stage of training, when a model learns from a huge collection of text, images or video. The model has 2.8 trillion parameters, the numbers a model tunes as it learns, against 1 trillion for MAI-Base-1, the starting model behind MAI-Thinking-1, which Microsoft announced in June. Microsoft says it built MAI-Base-2 end to end itself, preparing all of its training data in-house, and did not train it on answers produced by other AI models.
MAI-Base-2 is a mixture-of-experts model, a design that switches on only part of the model for each piece of text, so it runs faster than its full size suggests: 80 billion of its parameters do the work for each piece of text. Microsoft says it trained the model on 16,000 GB200 AI chips over 38 days in August and September, and that unlike MAI-Base-1 it learned from images as well as text.
Microsoft says MAI-Base-2 beats all four open-weight models it tested, meaning models whose trained files are published so anyone can download them and run them on their own computers. Its tests measure how well each model predicts writing it has not seen: the code for Microsoft AI's own software, plus sets of general questions, math and science it paid outside firms to write, to make it less likely that any model had seen them in training. Microsoft also estimates that its training method would need 4 to 12 times less computing power to match DeepSeek-V4-Pro-Base, which it calls the strongest free-to-download model that has had only this first stage of training.
At a town hall for employees on Thursday, Microsoft AI chief Mustafa Suleyman went further, according to a news report. "We now have a stronger pre-trained base model than the model that currently powers GPT 5.6 Sol," he said, adding that Microsoft used a quarter of the computing power OpenAI used to train its model. He said Microsoft compared the two privately, using technical information it has about OpenAI's models. Microsoft has not published that comparison, so outsiders cannot check it. Microsoft has invested billions of dollars in OpenAI and has used OpenAI's models in its own products, such as Copilot, according to the same report. Microsoft declined to comment, and OpenAI did not respond to a request for comment.
All of these results are Microsoft's own, on tests it chose, and three of the rivals were finished models rather than ones fresh from their first stage of training, which Microsoft says may affect the comparison. MAI-Base-2 is still going through further training before it is ready for use. Microsoft says it plans to use its models first in its own products and then offer them through Foundry, its service for businesses, and that building them entirely from its own data will let it license the models' trained files to trusted partners there. It has not said when that will happen, how the finished model will compare with the leading chatbots, or what it will cost.