Microsoft launches seven in-house AI models for image, voice, code and reasoning

Microsoft unveiled seven new in-house AI models at Microsoft Build 2026, developed entirely by its MAI team without distillation from third-party models or reliance on opaque data. The family spans five categories — image, transcription, voice, reasoning and code — and is designed to integrate into existing Microsoft products while also being available to external developers through platforms such as OpenRouter, Fireworks and Baseten.

The reasoning model MAI-Thinking-1 is the flagship of the group. It is a medium-sized model — 35 billion active parameters — with a 256,000-token context window that, according to internal evaluations with independent human raters, is preferred over Claude Sonnet 4.6 in blind comparisons. On the SWE-Bench Pro index, which measures performance on real-world software engineering tasks, it scores 52.8%, placing it on par with Claude Opus 4.6. On the AIME 2025 mathematical reasoning test it reaches 97%.

In image generation and editing, MAI-Image-2.5 ranks second on the Arena image editing leaderboard with a score of 1,403. The model includes a Flash variant designed for large-scale production at lower computational cost. It is already available in PowerPoint and is being rolled out to OneDrive.

MAI-Transcribe-1.5 is the family's speech-to-text model. According to tests using the public FLEURS benchmark, it outperforms equivalent models from OpenAI and Google in average accuracy across 43 languages, while operating up to five times faster than competitors as measured by the Artificial Analysis speed index. It integrates into Copilot, Teams, GitHub and Dynamics 365.

MAI-Voice-2 covers speech synthesis in 15 languages with the ability to adapt to a speaker's voice from a brief sample. It includes technical safeguards against unauthorised cloning and watermarks embedded in generated audio. A Flash variant, announced for the coming weeks, will reduce latency for real-time applications.

For software development, MAI-Code-1-Flash is a coding-specialist model with 5 billion active parameters that scores 51% on SWE-Bench Pro. Priced below Claude Haiku 4.5, it is integrated as one of the default models in VS Code and GitHub Copilot.

Alongside the model launches, Microsoft announces a collaboration with Mayo Clinic to jointly develop a clinical AI model. It will be trained on de-identified Mayo Clinic data and deployed first within its hospital system before being made available to other organisations through Microsoft Foundry. Mayo Clinic will retain ownership of the resulting model.

All models are accompanied by Microsoft Frontier Tuning, a service that allows organisations to fine-tune MAI models on their own workflows using reinforcement learning. The knowledge generated remains exclusively with each client and is not shared with other users. According to the company, a model fine-tuned internally for Excel tasks performs comparably to GPT-5.4 at up to ten times lower cost per generated token. McKinsey is among the first clients of the service, with similar efficiency gains reported against GPT-5.5.

Key points

• Microsoft launches seven in-house AI models without third-party distillation: image, transcription, voice, reasoning and code.
• MAI-Thinking-1 outperforms Claude Sonnet 4.6 in blind evaluations and scores 52.8% on SWE-Bench Pro.
• MAI-Transcribe-1.5 leads in accuracy across 43 languages and is up to five times faster than competing models.
• MAI-Image-2.5 ranks second on the Arena image editing leaderboard.
• MAI-Code-1-Flash scores 51% on SWE-Bench Pro with just 5 billion parameters, at a lower cost than Haiku 4.5.
• MAI-Voice-2 synthesises speech in 15 languages and can adapt to a voice from a brief sample.
• Microsoft Frontier Tuning lets companies fine-tune MAI models on their own data; the result belongs exclusively to the client.
• Microsoft and Mayo Clinic will jointly develop a clinical AI model, with Mayo Clinic retaining ownership.