Models

Microsoft Unveils Seven In-House MAI Models at Build 2026, Led by MAI-Thinking-1

On June 2, 2026, at Microsoft Build in San Francisco, Microsoft AI announced a family of seven in-house models spanning reasoning, coding, image, voice, and transcription. Flagship MAI-Thinking-1 is a mid-sized 35B-active-parameter MoE with a 256K context window, trained from scratch without third-party distillation; Microsoft says blind raters prefer it to Claude Sonnet 4.6 and that it matches Opus 4.6 on SWE-Bench Pro coding. Companion models include MAI-Code-1-Flash for Copilot/VS Code, MAI-Image-2.5, MAI Transcribe 1.5, and MAI-Voice-2.

Tech Insights Reporter Jun 2, 2026 7 min read San Francisco, CA

TLDR

Microsoft used the opening day of Build 2026 (June 2, 2026) to launch seven new Microsoft AI (MAI) models developed in-house—positioned as a multi-modal family for reasoning, coding, image, voice, and transcription. The centerpiece is MAI-Thinking-1, Microsoft’s first reasoning model: roughly 35 billion active parameters (MoE, ~1T total), 256K context, trained from scratch on clean commercially licensed data with no third-party distillation. Microsoft reports blind preference over Sonnet 4.6 and parity with Opus 4.6 on SWE-Bench Pro, with private preview availability on Azure AI Foundry.

The MAI model family

According to Microsoft’s official Build blog and Microsoft AI announcements:

MAI-Thinking-1: First reasoning model; mid-sized MoE (~35B active / ~1T total); 256K context; multi-step instructions, long-context reasoning, code generation; Foundry private preview.
MAI-Code-1 / MAI-Code-1-Flash: Inference-efficient agentic coding models tuned for GitHub Copilot and VS Code; Flash variant described as ~5B active parameters, positioned as cheaper than peer lightweight coding models.
MAI-Image-2.5 (+ flash): Text-to-image and image-to-image; Microsoft cites Arena AI leaderboard placements (#3 T2I, #2 I2I, surpassing Nano Banana 2 on editing); live in PowerPoint, rolling to OneDrive, on Foundry.
MAI Transcribe 1.5: Transcription accuracy claims across 43 languages; streaming coming.
MAI-Voice-2 (+ flash): Expanded language coverage (15+ additional languages) and new voice options.

Microsoft also said MAI models will be available via third-party hosts including Fireworks AI, Baseten, and OpenRouter, with Fireworks generally available on Foundry for governed multi-model access.

Build platform context

The model launch sat inside a broader Build platform story: Microsoft IQ context layers (Work IQ, Fabric IQ, Foundry IQ, new Web IQ), Microsoft Scout personal agent for Frontier customers, Agent 365 for local agents, Frontier Tuning private preview, hosted agents in Foundry, GitHub Copilot desktop app preview, and Surface RTX Spark Dev Box (up to 1 petaflop AI, 128GB unified memory). The MAI family is the clearest signal that Microsoft is shipping first-party frontier-class models to reduce exclusive dependence on external labs for high-volume developer and Office workloads.

Why this story matters

Build 2026 marks Microsoft’s most explicit first-party model push yet—framing MAI as a cost-efficient, enterprise-licensed alternative to Claude and GPT for coding and reasoning while keeping Azure multi-model. For developers and CIOs, it changes the default choice set inside Copilot, Foundry, and VS Code; for the AI market, it compresses margins and accelerates competition on unit economics as hyperscalers productize their own mid-weight frontier stacks.

Sources

Microsoft Official Blog: “Microsoft Build 2026: Be yourself at work” (blogs.microsoft.com, June 2, 2026) — primary Build overview with MAI specs and benchmarks.
Microsoft AI: “Building a hill-climbing machine: Launching seven new MAI models” and “Introducing MAI-Thinking-1” / “Introducing MAI-Code-1-Flash” (microsoft.ai, June 2, 2026).
CNBC and other same-day coverage of MAI-Code-1-Flash and reduced reliance on OpenAI (June 2–3, 2026).
Build 2026 news hub (news.microsoft.com/build-2026).

Featured Image Alt Text

Microsoft Build 2026 stage branding with MAI model family icons for thinking, code, image, voice, and transcription.