Microsoft completes pre-training of 2.8-trillion-parameter MAI-Base-2

What Microsoft built

Microsoft trained MAI-Base-2 on 16,000 Nvidia GB200 GPUs over 38 days in August and September. The run processed 50 trillion tokens of text and visual data. The model holds 2.8 trillion parameters in total but activates 80 billion for each input, a mixture-of-experts design that lets it keep a large model without running every parameter for every operation.

Microsoft expanded unique training data by about 36% compared with MAI-Base-1, which it introduced earlier this year. Images made up 18% of the new training mixture, part of what Microsoft calls native vision training for chart interpretation, document analysis and spatial reasoning. Microsoft says the model learned from organic training data rather than by distilling knowledge from other AI models.

Efficiency claims and their limits

Microsoft estimates its training approach is 4.2 to 12.3 times more compute-efficient than DeepSeek-V4-Pro-Base when matching performance on selected tests. The company says MAI-Base-2 outperformed the open-weight models included in its internal evaluations. TechRepublic noted that the figures estimate the compute needed to reach comparable performance, rather than showing that MAI-Base-2 is universally superior.

The comparisons rely on held-out evaluations Microsoft chose, and the company has not published equivalent comparisons against every leading proprietary model. Business Insider reported that Mustafa Suleyman, Microsoft's AI chief, told employees the company had privately compared MAI-Base-2 with OpenAI's models. Microsoft has not publicly released those results for independent assessment.

Licensing, not open release

Microsoft plans to license MAI-Base-2's weights to trusted partners through Microsoft Foundry rather than release them publicly. Microsoft has not announced pricing or general availability. TechRepublic noted that access conditions and customization costs will influence whether companies choose Microsoft's offering over genuinely open-weight alternatives.

The model is still undergoing post-training. The report added that businesses should not expect immediate access to MAI-Base-2 through Copilot or other Microsoft services. Microsoft says preparations for MAI-Base-3 are already underway.

TechRepublic reported that the model's text and vision capabilities could eventually support enterprise applications involving documents, spreadsheets and charts, but Microsoft has not demonstrated how those capabilities will translate into commercial products.

Why Microsoft is building its own models

Microsoft has invested roughly $13 billion in OpenAI and continues to use OpenAI models for more demanding tasks, eWeek reported. The company has also begun using its own MAI models in some products to lower costs. Suleyman said MAI models are deployed in more than half of Microsoft's products, according to Bloomberg, and that the goal is to reduce and ultimately eliminate what Microsoft pays Anthropic.

Microsoft amended its OpenAI agreement in April, ending exclusive access to OpenAI's intellectual property and eliminating Microsoft's payments to OpenAI, while OpenAI continued paying Microsoft with a cap, Yahoo Finance reported. Microsoft still supplies the computing infrastructure for OpenAI's models, eWeek reported.