Microsoft expands model families and agentic AI

At its Build 2026 developer conference, Microsoft announced an expansion of its AI model families and agentic-AI infrastructure and made Microsoft IQ, its unified intelligence layer, generally available, per SiliconANGLE and Reuters. SiliconANGLE describes Microsoft IQ as a four-tier stack, Work IQ, Fabric IQ, Foundry IQ, and Web IQ, intended to make agents and Microsoft Copilot context-aware across enterprise data and the web. According to SiliconANGLE, Work IQ APIs, which give agents direct access to signals inside Microsoft 365 such as people, emails, documents, and meetings, will become available on June 16, and a Fabric-hosted semantic layer will act as an "ontology" for structured business data.

SiliconANGLE reports that the Microsoft Superintelligence Team released a family of seven new in-house models, led by the company's first reasoning model, MAI-Thinking-1. Microsoft and SiliconANGLE describe it as a 35-billion-active-parameter model with a 128,000-token context window, trained from scratch with no distillation on commercially licensed data, positioning it as independent of OpenAI's outputs. Microsoft says blind-test raters preferred MAI-Thinking-1 to Anthropic's Sonnet 4.6 and that it matched Opus 4.6 on coding in the SWE-bench Pro benchmark (scoring about 53 percent), with strong results on math-reasoning benchmarks; the model is available in private preview on Microsoft's AI Foundry. These are vendor-reported figures that independent benchmarking has not yet confirmed.

SiliconANGLE also reports a model-agnostic, Model Context Protocol-native web-grounding capability that Microsoft claims returns relevant information blocks about 2.5 times faster than the next-best alternative. The combination of long-context models, an enterprise semantic layer, and standardized agent connectors reflects Microsoft's pitch that grounding and context, rather than raw model size, are the binding constraints for useful agents.

Reuters reported that Microsoft revealed prototype hardware under the name Project Solara, a family of devices designed to host AI agents rather than run a traditional operating system and apps. Reuters said the prototypes, roughly the size of a smart speaker or badge, are built on chips from Qualcomm and MediaTek, with demonstrated use cases such as documenting a medical visit, and quoted CEO Satya Nadella framing the devices as not just a new platform but "a set of new platform rules" that expand what developers can build.

As a general industry pattern, vendors building agent platforms have converged on three threads visible here: longer-context models, retrieval and grounding layers positioned as the main defense against hallucination, and APIs that expose proprietary enterprise signals to agents. Training a frontier-class model from scratch also fits a broader move by large platforms to reduce dependence on rival labs and control their own model supply chain.

• •Whether Work IQ APIs ship on June 16 as reported, and on what licensing and SLA terms.
• •Independent benchmark evaluations that confirm or dispute Microsoft's MAI-Thinking-1 comparisons to Sonnet 4.6 and Opus 4.6.
• •Project Solara form factors, chip partners, and the developer tooling required to run agents on-device.