Microsoft unveils seven new MAI AI models at Build 2026
The new lineup spans coding, cybersecurity, image generation and voice, extending the in-house model family Microsoft introduced at Build
Microsoft has announced six new AI models at its Ignite event, covering coding, cybersecurity, image generation and voice.
What Microsoft announced
The six models fall under the Microsoft AI brand, better known as MAI. They target four broad categories of work: writing code, defending systems, generating images and producing voice.
Cybersecurity stands out as the newest category. Microsoft’s earlier MAI lineup focused on reasoning, coding, images, transcription and voice. Security tools now join that list.
How the MAI family got here
The Ignite announcement builds on a larger reveal earlier in the year. At its Build 2026 developer conference, held on June 2, 2026, in San Francisco, Microsoft introduced a suite of seven MAI models.
That batch included MAI-Thinking-1, MAI-Code-1-Flash, MAI-Image-2.5, MAI-Transcribe-1.5 and MAI-Voice-2. Several came with lighter Flash variants, which typically trade some capability for speed and lower cost.
MAI-Thinking-1 was the headliner. Microsoft described it as its first advanced reasoning model, built from scratch on clean, commercially licensed data without relying on third-party models such as OpenAI’s.
AI, tech, and the markets they move—in one daily briefing.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
The specs were substantial. MAI-Thinking-1 runs with 35 billion active parameters and a context length of 256K.
It scored 97% on the AIME 2025 benchmark, a set of hard competition math problems. Microsoft also said it performed favorably against Anthropic’s Claude Sonnet 4.6 in blind evaluations.
The other Build models covered more practical ground:
- MAI-Code-1-Flash uses 5 billion parameters and is aimed at agentic coding workflows, where an AI carries out multi-step programming tasks rather than just suggesting a line.
- MAI-Image-2.5 and its Flash variant rank at or near the top of image generation leaderboards, and reportedly beat Google’s Nano Banana models on editing tasks.
- MAI-Transcribe-1.5 handles transcription in 43 languages with what Microsoft calls state-of-the-art accuracy.
- MAI-Voice-2 offers voice cloning in 15 languages and also comes in a Flash variant.
Mustafa Suleyman, Microsoft’s AI chief, pitched the lineup as competitive with prominent systems from Anthropic’s Claude series and Google’s image models.
Microsoft designed the models around efficiency, enterprise-grade data practices and smooth integration with tools it already owns. Those include GitHub Copilot, Visual Studio Code and Microsoft Foundry.
Ignite 2026 runs from November 17 to 20, 2026, with sessions dedicated to MAI on the schedule.
Why the in-house push matters
Microsoft has emphasized that its MAI models are trained on clean data and built without third-party distillation. Distillation is the practice of training a smaller model by copying the outputs of a larger one. Avoiding it means Microsoft can claim the models as genuinely its own.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.

