Microsoft used Build 2026 to announce seven in-house MAI models, including MAI-Thinking-1, a reasoning model described by Microsoft AI as a 35B-active-parameter mixture-of-experts system with a 256K context window.
Microsoft announced seven in-house MAI models at Build 2026, including a new reasoning model called MAI-Thinking-1.
In its official Build 2026 blog post, titled “Be yourself at work,” Microsoft said it had announced seven in-house MAI models. The company’s Build 2026 news hub also points readers to the keynote, the same official blog post, and related announcements across developer tools, Windows, Microsoft AI, and quantum computing.
The most detailed model highlighted in the provided source material is MAI-Thinking-1. Microsoft AI described it in a June 2, 2026 announcement as a reasoning model “trained from the ground up” on what it called enterprise-grade, clean, commercially licensed data. Microsoft AI also said the model was trained without third-party model distillation.
According to Microsoft’s Build 2026 blog excerpt, MAI-Thinking-1 is a mixture-of-experts reasoning model with 35 billion active parameters and a 256K context window. Microsoft positioned the model around low token cost and said access would be available through a Foundry private preview.
Microsoft also made benchmark claims comparing MAI-Thinking-1 with Anthropic’s Sonnet 4.6 and Opus 4.6 models. The provided Microsoft excerpt does not include the full benchmark methodology or detailed scores, so those comparisons should be treated as Microsoft’s own claims rather than independent evaluations.
The emphasis on active parameters suggests Microsoft is highlighting efficiency as part of the model’s design. Mixture-of-experts systems typically activate only part of a model for a given request, but the sources provided do not give further architectural details beyond the 35B-active-parameter description.
Microsoft’s Build 2026 news hub frames the announcement as part of a broader developer event covering several areas of the company’s product strategy. The hub links the keynote and featured Build announcements across Developers, Windows, Microsoft AI, and Quantum.
Within that broader context, the MAI model announcements indicate that Microsoft is continuing to build its own model portfolio rather than relying only on partner models. The Microsoft AI announcement for MAI-Thinking-1 specifically stresses that the model was trained internally from the ground up and without third-party model distillation.
The provided Microsoft materials establish the headline specifications for MAI-Thinking-1 and confirm that seven in-house MAI models were announced. They do not, however, provide full details for all seven models in the excerpted sources.
The materials also do not provide public availability dates beyond the reference to Foundry private preview access. Developers and enterprise customers will likely need Microsoft’s full Build 2026 documentation or Foundry preview information to understand pricing, availability, supported regions, safety documentation, and deployment options.
For now, the main confirmed facts are that Microsoft announced seven in-house MAI models at Build 2026, identified MAI-Thinking-1 as a 35B-active-parameter mixture-of-experts reasoning model with a 256K context window, and presented its benchmark comparisons and cost positioning as part of Microsoft’s own launch messaging.
Microsoft announced seven in house MAI models at Build 2026, including a new reasoning model called MAI Thinking 1.
Microsoft expands its MAI model lineup In its official Build 2026 blog post, titled “Be yourself at work,” Microsoft said it had announced seven in house MAI models.
The company’s Build 2026 news hub also points readers to the keynote, the same official blog post, and related announcements across developer tools, Windows, Microsoft AI, and quantum computing.
Continue reading
