Microsoft used its Build 2026 conference on June 2 to launch seven new in-house AI models under its MAI (Microsoft AI) brand, covering reasoning, coding, image generation, voice and transcription. The headline releases are MAI-Thinking-1, which Microsoft presents as its first reasoning model, and MAI-Code-1-Flash, a small agentic coding model that is rolling out inside GitHub Copilot. For developers, the launch adds a new first-party model family to Microsoft Foundry and signals that Microsoft intends to compete with its own models, not only host those of its partners.

Key Facts#

  • Announcement: seven MAI models unveiled at Microsoft Build on June 2, 2026, spanning reasoning, coding, image, voice and transcription.
  • MAI-Thinking-1: described by Microsoft as its first reasoning model. Launch coverage reported a mid-sized design with 35 billion active parameters.
  • Context window: Microsoft Foundry documentation lists a 256K-token budget per request, with output capped at 64K tokens.
  • Foundry status: MAI-Thinking-1 is in preview in Foundry, available as a Global Standard deployment only, without provisioned throughput (PTU) support.
  • MAI-Code-1-Flash: a small, inference-efficient agentic coding model that GitHub said is rolling out in Copilot, starting with VS Code.
  • Other platforms: according to launch coverage, Microsoft also confirmed availability through Fireworks AI, Baseten and OpenRouter.

What Happened#

Microsoft AI, the division that builds Microsoft's own models, published the launch under the title "Building a hill-climbing machine" and also released a transcript of its MAI keynote at Build. According to reporting on the announcement, Microsoft said MAI-Thinking-1 was trained from scratch on commercially licensed data instead of being distilled from another company's models, and that it was designed for efficiency and low per-token cost rather than maximum size.

Foundry's documentation describes MAI-Thinking-1 as a reasoning model for workloads that depend on multi-step reasoning, such as math, coding and enterprise analysis. It is served through a chat completions endpoint under a /mai/v1 path on a Foundry resource, following OpenAI SDK conventions, and it can return encrypted reasoning content that clients pass back unchanged on later turns so the model keeps its reasoning state across a conversation.

The coding model shipped where developers already work. In its Build recap, GitHub said MAI-Code-1-Flash, Microsoft's latest small-tier coding model, was rolling out in GitHub Copilot starting with VS Code, and it described the model's quality as best in class for its size compared with other lightweight models. That is a vendor claim that independent benchmarks will need to confirm. Microsoft also released new versions of its image generation, voice and transcription models as part of the seven-model launch.

Background#

Microsoft has spent years as one of OpenAI's biggest financial backers and has built much of its Copilot portfolio on OpenAI models. At the same time, it has been assembling its own model capability. Its earlier voice model, MAI-Voice-1, a neural text-to-speech model built on Microsoft's in-house speech foundation models, is available in preview through Azure Speech, and the Build launch extended that effort to reasoning and coding.

The relationship with OpenAI also changed shortly before Build. In April 2026, the two companies amended their partnership so that OpenAI can serve its products on any cloud, while Microsoft kept a license to OpenAI's models and products through 2032 (see the amended Microsoft–OpenAI partnership). With exclusivity loosened on both sides, a credible first-party model family gives Microsoft more control over cost, supply and product direction.

Why It Matters for Developers#

For .NET developers on Azure, the MAI family is another set of options in the Foundry catalog, alongside models from OpenAI, Anthropic and others. The practical questions are the usual ones: quality on your tasks, latency, price and operational constraints.

  • Code against abstractions. Because MAI-Thinking-1 follows OpenAI-style chat completions conventions, applications built on IChatClient from Microsoft.Extensions.AI can evaluate it with configuration changes rather than rewrites. Our guide to Azure OpenAI and Microsoft Foundry for .NET covers endpoints, deployments and authentication.
  • Check the preview constraints. Global Standard deployment means requests can be processed in any Foundry region, which matters for data residency, and the lack of PTU support means no reserved capacity for now. The 64K output cap also limits very long generations.
  • Handle reasoning state carefully. If you use encrypted reasoning content, store it with the conversation and send it back unchanged. Treat it as opaque data in your persistence layer.
  • Evaluate before switching. Vendor claims about quality per parameter are a starting point, not a decision. Run your own test sets with the tools described in evaluating AI applications in .NET.

In GitHub Copilot, MAI-Code-1-Flash adds a small, fast option for everyday edits and explanations. Microsoft's training material for its successor, MAI-Code-1.1-Flash, lists .NET among its supported languages and stresses that generated code is a proposed implementation that still needs human review. That advice applies to every coding model, as discussed in AI-assisted .NET development.

What's Next#

The MAI family has kept moving since Build. Microsoft's current Learn training material lists newer versions such as MAI-Code-1.1-Flash, MAI-Voice-2, MAI-Transcribe-1.5 and an MAI-Image-2.5 family, along with a security-focused MAI-Cyber-1-Flash model, while Foundry's documentation still labels MAI-Thinking-1 as a preview. The open questions are whether Microsoft will add provisioned throughput and regional deployments for enterprise workloads, how MAI models will be priced against partner models in Foundry, and how prominently Microsoft will use its own models inside Copilot products.

Sources#