TL;DR
- Foundry Rollout: Microsoft has added seven in-house MAI models across reasoning, code, image, voice, and transcription workflows.
- Reasoning Model: MAI-Thinking-1 is in private preview with 35 billion active parameters and a 256K-token context window.
- Customer Control: Microsoft says developers will be able to tune model weights, a deeper option than prompt engineering alone.
- Competitive Context: Google and Anthropic have recently moved Gemini and Claude models into similar developer-focused workflows.
Microsoft has launched its in-house MAI model family into Foundry across reasoning and multimodal workloads, turning the Build 2026 update into a broader developer push rather than a single-model release. MAI-Thinking-1 is the flagship reasoning system, while additional code, image, and transcription models extend the lineup into everyday developer and enterprise workflows.
Developers and enterprise teams get limited Foundry access first as Microsoft moves more first-party systems toward customer testing. Foundry remains Microsoft’s platform for finding, deploying, and governing AI models, but the wider MAI rollout puts Microsoft-owned models into the same decision path where customers already compare OpenAI, Anthropic, Google, and specialist speech or coding tools.
MAI-Thinking-1 Leads the Foundry Push
Microsoft’s Foundry rollout covers AI systems for Microsoft Foundry across reasoning, image, voice, and speech. Microsoft AI also launched seven new MAI models across image, voice, transcription, coding, and reasoning, with MAI-Thinking-1 in private preview for Foundry users and a MAI Playground public preview planned later.
MAI-Thinking-1 carries the central technical role in the rollout. A sparse Mixture-of-Experts design routes each task to selected expert subnetworks rather than activating the whole system, which can preserve overall model capacity while limiting the compute used for a given request.
Its specification lists 35 billion active parameters, roughly 1 trillion total parameters, a 256K-token context window, function calling, developer instructions, and compatibility with the Chat Completions API. A 256K-token window gives the model room to consider longer codebases, documents, or instructions in one prompt, while Chat Completions compatibility reduces integration work for teams that already use that API pattern.
MAI-Thinking-1 is Microsoft’s flagship reasoning model for the Foundry push. Kyle Daigle, Microsoft Developer CMO and COO of GitHub, positioned MAI-Thinking-1 for complex chained tasks, long-context reasoning, and code generation.
Microsoft claims the model was trained on commercially licensed data without distillation from third-party models. That posture is aimed at enterprise customers that weigh intellectual-property risk, data provenance, and vendor control alongside benchmark scores.
Microsoft AI CEO Mustafa Suleyman tied the launch to deeper developer control over first-party models.
“For the first time developers will be able to tune the weights of the model themselves.”
Mustafa Suleyman, Microsoft AI CEO (via Microsoft AI)
Weight tuning gives an enterprise a deeper adaptation path than prompt engineering or retrieval layers alone. Microsoft’s evaluation materials put MAI-Thinking-1 at software-engineering benchmark parity with leading models and Sonnet 4.6 in blind side-by-side evaluations. Customer trials will determine whether those benchmark comparisons turn into reliable production gains.
Seven Models Stretch Across Code, Image, Voice, and Speech
Microsoft’s broader MAI family keeps the launch from becoming only a reasoning-model release. MAI-Code-1-Flash is a 5 billion parameter coding model integrated into GitHub Copilot, VS Code, and the Microsoft developer stack, while MAI-Code-1 is available in GitHub Copilot and VS Code.
MAI-Image-2.5 and its Flash variant add text-to-image generation and image editing with control-with-preservation features. MAI-Image-2.5 also reached No. 3 for image generation model families, extending the model’s Arena ranking into a wider first-party image push.
MAI-Voice-2 covers voice cloning and voice prompting across more than 15 languages. MAI-Transcribe-1.5 supports 43 languages with domain-specific terminology support and a Microsoft-presented five-times-faster transcription claim.
For organizations that already split AI work across code assistants, image systems, and speech pipelines, the MAI lineup turns Foundry into a place to evaluate multiple Microsoft-owned model classes under one governance layer.
Microsoft’s models are also planned for wider developer availability on OpenRouter, Fireworks, and Baseten, alongside Foundry and first-party products. For teams that mix providers by cost, governance, latency, and workload fit, the added route makes availability an operational choice rather than a purely Microsoft-platform decision.
Google and Anthropic Add Competitive Pressure
Google and Anthropic give the timing a sharper competitive edge. Gemini 3.5 Flash became generally available on May 19 for developers through Google AI Studio, Android Studio, and enterprise channels, with agentic workflows, coding, long-horizon tasks, and multimodal understanding in focus.
Anthropic launched Claude Opus 4.8 on May 28 with benchmark and agentic-workflow improvements over Opus 4.7. Microsoft had already released three in-house AI models through Foundry and MAI Playground in an earlier rollout; the latest lineup widens that first-party track across reasoning, code, image, voice, and transcription.
MAI Playground is Microsoft’s next concrete availability channel for MAI-Thinking-1. Until that channel opens or Foundry access expands, Microsoft keeps direct testing of the reasoning model tied to the current program, while customers judge whether first-party tuning, data posture, and distribution breadth justify adding another model family to production evaluations.
