Details
- Microsoft AI says it is launching seven new models at Build spanning reasoning, code, image, transcription, and voice.
- The lineup includes MAI-Thinking-1, described as the company’s flagship reasoning model, trained from scratch on a clean data lineage.
- MAI-Code-1-Flash is positioned for coding and is rolling out in GitHub Copilot, with Microsoft saying it beats Claude Haiku 4.5 on its tested coding benchmarks while using up to 60% fewer tokens.
- MAI-Image-2.5 and MAI-Image-2.5-Flash target image editing and text-to-image work, with controllable edits designed to preserve faces, logos, and fine details.
- MAI-Transcribe-1.5 focuses on speech recognition, with Microsoft claiming strong multilingual accuracy and speed performance across 43 languages.
- MAI-Voice-2 and MAI-Voice-2-Flash add expressive text-to-speech features, including emotion control, stable speaker identity, and code-switching.
- Microsoft says the models were built from scratch, are designed to work as a family, and are available through Microsoft Foundry.
Impact
Microsoft is broadening its AI stack from hosted model access toward a fuller first-party model family that can power coding, voice, transcription, and image products inside its own ecosystem. The GitHub Copilot rollout gives it an immediate enterprise channel, while the “from scratch” framing signals a push for tighter control over efficiency and model behavior. The move increases pressure on OpenAI, Anthropic, and Google by making Microsoft both a platform partner and a model supplier in key workflows.
