Microsoft expands AI model family and agent infrastructure

Microsoft Corp. announced a broad expansion of its AI model family and agent infrastructure at its Build developer conference in San Francisco, saying the updates are meant to give developers and business users more data and context when they deploy AI systems.

The centerpiece of the announcement was Microsoft IQ, which the company said is now generally available. Microsoft described it as a unified intelligence layer for AI agents and Microsoft Copilot that makes them context-aware and personalized for organizations. Instead of acting like a generic chatbot, the system connects foundation models to a company’s data and business logic in an effort to reduce hallucinations, according to Microsoft.

Microsoft said Microsoft IQ includes several layers. Work IQ is designed to capture how people work inside Microsoft 365 by connecting organizational systems and external sources, including people, emails, documents and meetings, and by mapping how those items relate to one another. The company said Work IQ application programming interfaces, or APIs, will be available on June 16 so agents can access data directly.

Fabric IQ, which runs on Microsoft Fabric, provides a semantic data foundation that Microsoft described as an “ontology” for organizing structured business data. Foundry IQ is the layer that ties the system together by retrieving information from unstructured sources such as wikis, policies, contracts and the live web.

Microsoft also introduced Web IQ, the newest member of the family. The company said Web IQ gives agents fast grounding in real-world information by using web search and that it is model agnostic and native to the Model Context Protocol, or MCP, which is a standard for connecting models to tools and data. Microsoft said it returns relevant information blocks nearly two and a half times faster than the next best alternative.

The company paired the infrastructure announcement with a set of new in-house models from its Microsoft Superintelligence Team. That group released seven models, led by MAI-Thinking-1, Microsoft’s first reasoning model. Reasoning models use chain-of-thought style processing before producing answers. Microsoft said MAI-Thinking-1 has 35 billion active parameters, a 128,000-token context window, and is tuned for efficiency and performance at low token cost.

Microsoft said independent raters preferred MAI-Thinking-1 to Anthropic’s Sonnet 4.6 in a blind test, and that it can match Opus 4.6 on coding tasks in SWE Bench Pro. The model is available now in private preview on Microsoft AI Foundry, along with a flash version built for speed and efficiency.

The company also announced MAI-Image-2.5 and a flash variant. These models support text-to-image generation and image-to-image editing, which lets users provide sketches or other visuals as part of the prompt. Microsoft said the new image model is live in PowerPoint and rolling out on OpenDrive.

Other additions include MAI-Transcribe-1.5 for speech recognition across 43 languages, with streaming support coming soon, MAI-Voice-2 and its flash variant in more than 15 additional languages, and MAI-Code-1, a coding model tuned for GitHub that is already in Copilot and VS Code.

For security, Microsoft launched Codename MDASH, a multimodel agentic-security system that uses more than 100 agents to look for exploitable bugs by reasoning about data flow, business logic and exploit chains. Microsoft said the system provides context-aware fixes inside the Developer Portal.

The company also introduced Scout, a personal agent for frontier customers that it said stays attentive at all times. Microsoft said Scout is built on OpenClaw and Work IQ, understands how a user works, uses tools on the user’s computer and in the organization such as Teams and Outlook, and can handle tasks like meeting prep, scheduling conflicts and routine work without needing to ask first. Microsoft said it will share more as Scout rolls out more broadly.