Seven MAI models for reasoning, programming, images, transcription, and voice make up Microsoft's new artificial intelligence family. The company announced it on June 2, 2026, presenting it as a set of internally developed models designed to be integrated into both its products and services for developers.
Microsoft AI's announcement lists MAI-Thinking-1, MAI-Code-1-Flash, MAI-Image-2.5 and its Flash variant, MAI-Transcribe-1.5, and MAI-Voice-2, along with an upcoming Flash voice variant. The company describes them as a multimodal family: distinct models built on shared infrastructure and intended to work within a single ecosystem.
A family of models for five tasks
MAI-Thinking-1 is the flagship reasoning model. Microsoft defines it as a medium-sized model and says it matches leading models on software engineering tests and demonstrates advanced mathematical reasoning capabilities. In its blind comparative evaluation, the company says participants preferred it to Sonnet 4.6; the announcement did not publish numerical results from that evaluation.
Microsoft's official blog specifies that MAI-Thinking-1 has 35 billion active parameters and a context window of 256,000 tokens. The company says it was trained from scratch, without distillation from third-party models, and is available in Microsoft Foundry in private preview. It also says the model matches Opus 4.6 in coding capabilities on SWE Bench Pro and was designed for complex, multi-step instructions, reasoning over long contexts, and code generation.
MAI-Code-1-Flash is an inference-efficient programming model. Microsoft describes it as an agentic model adapted for GitHub Copilot, Visual Studio Code, and the Microsoft stack; Microsoft AI's announcement attributes 5 billion active parameters to it and says it is comparable to Haiku at a lower cost. The Microsoft Build blog calls it MAI-Code-1 and says it is already available in Copilot and VS Code.
The visual offering consists of MAI-Image-2.5 and its Flash variant. Microsoft says they support text-to-image generation and image editing, and presents Flash as a more efficient option. In its Build post, the company places MAI-Image-2.5 third on Arena AI for text-to-image and second for image-to-image tasks, where it says the model outperforms Nano Banana 2. These rankings and comparisons are reported by Microsoft.
For audio, Microsoft announced MAI-Transcribe-1.5, which transcribes 43 languages and includes built-in support for domain-specific terminology. The company claims it offers state-of-the-art accuracy and is five times faster than competing models, without identifying which models it used for the comparison in that post. For voice, MAI-Voice-2 generates speech in 15 languages and can adapt a voice from a short sample; the MAI-Voice-2-Flash variant, described as more efficient and lower-cost, was announced as coming soon.
| Model | Described function | Reported detail |
|---|---|---|
| MAI-Thinking-1 | Reasoning and code | 35 billion active parameters; 256,000-token context. |
| MAI-Code-1-Flash | Agentic programming | 5 billion active parameters; integration with GitHub Copilot and VS Code. |
| MAI-Image-2.5 and Flash | Image generation and editing | Text-to-image and image-to-image; Flash is geared toward efficiency. |
| MAI-Transcribe-1.5 | Speech transcription | 43 languages and domain-specific terminology. |
| MAI-Voice-2 and Flash | Voice generation | 15 languages; voice adaptation from a short sample. |
Access through Foundry and external platforms
Microsoft plans to offer the models through Microsoft Foundry and its products, while expanding developer access through OpenRouter, Fireworks, and Baseten. Microsoft AI's post adds that, for the first time, developers will be able to tune the model weights themselves. The announcement does not specify when each model will be available on each platform or how much they will cost.
Availability details vary across Microsoft's posts. The Build blog says MAI-Thinking-1 is in private preview in Foundry and that MAI-Image-2.5 is already in PowerPoint, with rollout to OneDrive, as well as coming to Foundry. By contrast, the Foundry roundup published for Build includes four models in public preview, listing them as MAI-Thinking-1, MAI-Image-2.5, MAI-Transcribe-2, and MAI-Voice-2. That post calls the transcription model MAI-Transcribe-2, while Microsoft AI's announcement and the Build blog use MAI-Transcribe-1.5.
The naming also differs for the coding model: Microsoft AI's post calls it MAI-Code-1-Flash, while the Build blog presents it as MAI-Code-1. The pages do not explain whether these names refer to different variants or reflect a naming update. Microsoft offers no further clarification there.
Frontier Tuning and models tailored to each company
The other central part of the announcement is Microsoft Frontier Tuning, an approach to fine-tuning that uses reinforcement learning in real-world environments. Microsoft says organizations can train models on their own workflows and data within their own environment, while retaining ownership of the tuned model. The company describes these environments as training spaces accessible only to each organization.
Microsoft says its tuned models can be more efficient and perform better than general-purpose alternatives on specific tasks. As an example, it claims its model for Excel matches GPT 5.4 with up to ten times the efficiency. It also says that, when tuned to an enterprise organization's standards, MAI achieved the highest win rate among the models evaluated, at approximately one-tenth the cost. The announcement does not identify the organization or specify the testing protocol.
People – you – must always remain in control.
— Mustafa Suleyman
In the post, Mustafa Suleyman presents this development as part of a long-term strategy: Microsoft says it trains its reasoning models from scratch, does not distill models from other labs, and uses clean, traceable datasets with commercial licenses. It also says it designs its systems alongside its Maia 200 silicon and is already seeing a 1.4x efficiency improvement as a result of this work.
Mayo Clinic to participate in a healthcare model
Microsoft and Mayo Clinic announced a collaboration to create an AI model for healthcare. The project will combine Mayo Clinic's clinical expertise and de-identified clinical data with Microsoft's AI capabilities. The company says the system will be designed for clinical reasoning and a range of healthcare applications.
The first deployment is planned within Mayo Clinic's environment. Microsoft cites earlier and more accurate diagnoses, as well as treatment planning, among the potential goals. Once validated, the model would be offered to other organizations through Foundry. According to the announcement, Mayo Clinic will own the model; no launch timeline was announced.
Availability and cost will be decisive in assessing the family's practical reach, but Microsoft did not publish model pricing in the announcement. It did outline a distribution roadmap and various integrations; access and the status of each model depend on the channel, according to the company's posts.
Microsoft also said it expects to rapidly expand its compute capacity and capabilities over the next year. For developers, the next concrete milestones will be the availability of each model in Foundry and on external platforms, and the arrival of MAI-Voice-2-Flash, which the company described as coming soon.
Sources: DEV Community, Microsoft AI, The Official Microsoft Blog, Microsoft Foundry Blog
