Microsoft unveils MAI-Thinking-1, its first AI reasoning model, with MoE architecture and 35 billion active parameters. The firm is banking on compact, specialized and more efficient models to reduce its dependence on OpenAI and Anthropic.
Microsoft is starting to shed its historical dependencies on OpenAI and Anthropic. The Redmond firm presents this Tuesday, June 2, during Build, its annual developer conference, a family of models ofGenerative AI cutting edge. This new version is delivered by the Microsoft AI Superintelligence Team, the team dedicated to artificial intelligence launched in November 2024 and led by Mustafa Suleyman, co-founder of DeepMind. Seven specialized models are announced, Microsoft presents the five main ones.
A cutting-edge model…
Called MAI (for Microsoft AI), the family was designed from scratch by the company. Microsoft has focused its efforts on two differentiating assets: legal risk on the one hand, and compactness on the other. Everything was entirely trained on commercially usable quality data. The main model, MAI-Thinking-1, the firm’s first reasoning AI, arrives with an MoE architecture and only 35 billion active parameters for inference.
Microsoft is not yet revealing the details of the benchmarks, but the model would be on par with Claude Opus 4.6 on SWE-bench Pro, a benchmark test for code in agentic mode. The firm has not yet communicated its price list, while ensuring that the model would offer cutting-edge performance at a price much lower than that of Anthropic models. The only downside: a context of 128,000 tokens, at a time when the latest cutting-edge models manage at least a million. The model is already available in Microsoft Foundry.
…and specialized models
At the same time, Microsoft’s Superintelligence team unveils MAI-Image-2.5 and MAI-Image-2.5 Flash, an update to its cutting-edge text-to-image model positioned as a direct competitor to Gemini Nano Banana Pro. In benchmarks, it would be superior to the Google model. Designed for creative workflows, it generates images from a prompt as well as editing them. It’s now integrated with PowerPoint, OneDrive, and Foundry. Microsoft ensures that it offers the best quality/price ratio on the market.
The firm is also launching MAI-Transcribe-1.5, an update to its transcription model displaying SOTA scores on more than 43 languages. The team took the opportunity to launch MAI-Voice-2 and MAI-Voice-2 Flash, which generate voices from a prompt in the majority of current languages. Microsoft finally releases MAI-Code-1, a specialized model for rapid, very efficient code editing. It is available in Copilot and VS Code. MAI models will also be accessible natively on Fireworks AI, Baseten and OpenRouter.
Build’s other AI announcements
At the same time, Microsoft is announcing new agentic products. Starting with Microsoft Scout, a personal autonomous work agent (on the positioning of Claude Cowork) accessible today to a panel of testers. Developed on the basis of OpenClaw and merged with WorkIQ, Microsoft’s AI agent, it integrates with Teams and Outlook to proactively manage all digital workplace tasks. Microsoft also unveils its agent platform, Microsoft Agent Platform, which covers the entire agent lifecycle: from development in GitHub to deployment in Foundry, to exposure in M365 and Teams.
The firm adds Frontier Tuning, which applies reinforcement learning to adapt the underlying model of an agent to the company’s data and workflows, within its compliance scope (a rather vague notion, we agree). For these agents to be relevant, they must know the context of the company. That’s what Microsoft IQ is all about, a software layer that connects enterprise data to agents and natively interfaces with WorkIQ, Fabric IQ, and Foundry IQ. Microsoft is finally launching Web IQ, a web search engine designed for agents, MCP compatible and up to 2.5 times faster than its competitors.
With this first MAI family, Microsoft is reaching a milestone. The Redmond firm is advancing on Google’s territory: that of specialized and efficient models. It does not claim to compete head-on with the most powerful models on the market. Its bet is elsewhere: more compact models, designed for specific uses, at a price that should make the difference. The benefit is twofold. Microsoft is reducing its historical dependence on OpenAI and Anthropic, and thereby accomplishing the mission for which the Superintelligence Team was (unofficially) created. The challenge is not to win over new customers, but to better serve your own, by keeping them on your own models rather than referring them to the competition.
