Microsoft plans to license the weights of its 2.8T-parameter MAI-Base-2 through Foundry

The license is the part of the post a buyer can act on, and the post words it as a plan. Microsoft AI wrote: "Our pre-training is built on organic data which does not involve distillation from other models." ^[4] That, it said, "will enable us to license our model weights to trusted partners through Foundry" ^[1]. Microsoft did not publish a release date, license terms, or the hardware a licensee would need to serve a model of this size. Until it does, an enterprise buyer is weighing an intention.

Microsoft is already a first-party model maker. It announced MAI-Thinking-1 in June, post-trained on top of MAI-Base-1 ^[7]. Business Insider reports that Microsoft also offers OpenAI models to its Azure cloud customers and has used them in Copilot ^[12], and that the relationship with OpenAI has frayed over the past year or so as Microsoft has worked on models of its own ^[14]. What a weights license would add is a channel in which partners receive the model itself ^[1]. We do not yet know whether that makes Microsoft a supplier in revenue terms or only in intent.

The tradeoff is with OpenAI. Mustafa Suleyman, who runs Microsoft AI, told staff: "We now have a stronger pre-trained base model than the model that currently powers GPT 5.6 Sol," and added that Microsoft "used one-quarter of the compute that OpenAI used to train their model" ^[20]. Business Insider noted that Microsoft owns a stake in OpenAI that is likely worth more than $100 billion, and that criticism of OpenAI could dent the value of that investment ^[11]. Microsoft declined to comment and OpenAI did not respond ^[13]. In our view, Microsoft's interest as an OpenAI shareholder and its interest as a prospective licensor of a competing model now point in different directions.

A skeptic would say Microsoft is grading its own paper. Some of that holds. Microsoft says MAI-Base-2 achieves lower bits-per-byte than the other models it tested ^[10]. But the held-out code set is Microsoft's own codebase ^[8], and Kimi K3, Qwen3.8-2.4T-A95B and MiMo-V2.6-Pro-RL were available only as post-trained checkpoints, which Microsoft says may affect the results ^[9]. The comparison with OpenAI rests on technical information Microsoft has about OpenAI's models, Suleyman said ^[21]. Against that, the STEM, math and Q&A sets were commissioned from vendors to reduce contamination risk, and Microsoft says it decontaminated all evaluation sets against its training data ^[8].

Efficiency is the headline figure, but Microsoft did not take it as a saving. Business Insider wrote that pre-training takes massive amounts of expensive computing power, so Microsoft could save money if it has found a way to do it more efficiently ^[15]. The post says "we have put that saving into training a larger model" ^[18]. MAI-Base-2 has 2.8 trillion total parameters against 80 billion active, about 35 to 1 ^[17], and Microsoft describes it as larger and sparser than MAI-Base-1 ^[16].