Microsoft AI launches MAI model family for private enterprise workflow tuning

Jun 2, 2026

The Microsoft AI lab released seven MAI models at Build to anchor its new Humanist Superintelligence initiative. These models are built from scratch with zero distillation (training a smaller model to mimic a larger one) from third parties. The flagship MAI-Thinking-1 reasoning model matches Claude Sonnet 4.6 in human preference evaluations.

MAI-Code-1-Flash parameters: 5 billion
MAI-Code-1-Flash SWE-Bench Verified: 71.6
MAI-Transcribe-1.5 speed: 1 hour audio in under 15 seconds
MAI-Image-2.5-Flash input price: $1.75 per 1M tokens
MAI-Voice-2 language support: 15 languages

This launch signals a move toward self-sufficiency, co-designing models with Maia 200 silicon for a 1.4x efficiency boost. By avoiding distillation, Microsoft AI scales performance via its own compute and data. The MAI-Image-2.5 release already validates this shift, securing a top-three spot on the Arena image leaderboard for text-to-image generation.

Organizations can use Frontier Tuning to adapt models to workflows using reinforcement learning (learning through trial and error) in private environments. MAI-Code-1-Flash is rolling out in GitHub Copilot, while multimodal models are available via Microsoft Foundry. A specialized healthcare model co-created with the Mayo Clinic is also in development for clinical reasoning.

View the full update on microsoft.ai

Microsoft AI

@MicrosoftAI1d ago

Seven new models launching at Build: let’s go! Reasoning. Code. Image. Transcribe. Voice. Built from scratch on a clean data lineage, designed for efficiency, working seamlessly as a family of models Thread 🧵 #MSBuild https://t.co/g3WQIcIQ24

4813.1k

View on X

Still wondering? A few quick answers below.

Microsoft Frontier Tuning is a reinforcement learning framework that allows organizations to adapt MAI models to their specific workflows. By training on the trace of real work—such as sequences of decisions and actions—companies can build custom models that retain institutional knowledge within their own private, controlled environments.

MAI-Thinking-1 is a medium-sized reasoning model developed from scratch to compete with leading models in its weight class. In blind side-by-side evaluations, it reached human preference parity with Claude Sonnet 4.6. It is specifically optimized for STEM reasoning and coding tasks using a data-driven reinforcement learning framework.

Microsoft and the Mayo Clinic are co-creating a frontier AI model specifically for healthcare. This model combines clinical expertise and de-identified data to excel at clinical reasoning and treatment planning. It will be deployed within the Mayo Clinic environment before becoming available to other organizations via Azure Foundry.

Every HeadsUpAI update is written based on its original source and reviewed before it's published. Read our editorial standards →

Keep reading

Microsoft brings MAI reasoning models to Fireworks for enterprise fine-tuning

Microsoft AI is launching a family of seven in-house models and partnering with Fireworks AI to enable weight-level customization. This allows organizations to integrate proprietary institutional knowledge directly into Microsoft's frontier reasoning models.

Mistral AIMar 20

Mistral AI Launches Forge to Train Frontier Models on Proprietary Data

Mistral AI launched Forge, a system for enterprises to build frontier-grade AI models on their own internal knowledge rather than public data. Forge handles pre-training, post-training, and reinforcement learning, with ASML, Ericsson, and the European Space Agency among early partners.

Microsoft MAI-Transcribe-1.5 delivers top tier accuracy at 276x real time speed

Artificial AnalysisYesterday

Microsoft MAI-Transcribe-1.5 delivers top tier accuracy at 276x real time speed

Microsoft has released MAI-Transcribe-1.5, a speech-to-text model that ranks third for accuracy while processing audio at 276x real-time speed. The model leads the accuracy-speed Pareto frontier, offering a high-performance alternative for high-volume enterprise audio workloads.

OpenRouter adds Microsoft MAI models for high speed multimodal generation

OpenRouterYesterday

OpenRouter adds Microsoft MAI models for high speed multimodal generation

OpenRouter has integrated Microsoft’s new in-house MAI-Image-2.5, MAI-Transcribe-1.5, and MAI-Voice-2 models into its unified API. These models provide a high-performance stack for image, speech, and voice tasks built entirely without third-party distillation.

What is Microsoft Frontier Tuning?

How does MAI-Thinking-1 compare to other reasoning models?

What is the significance of the Mayo Clinic partnership?

Keep reading

Microsoft brings MAI reasoning models to Fireworks for enterprise fine-tuning

Microsoft brings MAI reasoning models to Fireworks for enterprise fine-tuning

Mistral AI Launches Forge to Train Frontier Models on Proprietary Data

Mistral AI Launches Forge to Train Frontier Models on Proprietary Data

Microsoft MAI-Transcribe-1.5 delivers top tier accuracy at 276x real time speed

Microsoft MAI-Transcribe-1.5 delivers top tier accuracy at 276x real time speed

OpenRouter adds Microsoft MAI models for high speed multimodal generation

OpenRouter adds Microsoft MAI models for high speed multimodal generation

Keep reading

Microsoft brings MAI reasoning models to Fireworks for enterprise fine-tuning

Microsoft brings MAI reasoning models to Fireworks for enterprise fine-tuning

Mistral AI Launches Forge to Train Frontier Models on Proprietary Data

Mistral AI Launches Forge to Train Frontier Models on Proprietary Data

Microsoft MAI-Transcribe-1.5 delivers top tier accuracy at 276x real time speed

Microsoft MAI-Transcribe-1.5 delivers top tier accuracy at 276x real time speed

OpenRouter adds Microsoft MAI models for high speed multimodal generation

OpenRouter adds Microsoft MAI models for high speed multimodal generation