Building a Hill-Climbing Machine: 7 New MAI Models Released

Microsoft AI announced a new lineup of 7 MAI model families, declaring the construction of a superintelligence lab that will define the next generation of AI. These models are designed to target a multimodal ecosystem spanning diverse domains including image, voice, transcription, coding, and reasoning.

The key model lineup is as follows:

• MAI-Thinking-1: Microsoft AI's flagship reasoning model, excelling at software engineering and mathematical reasoning. In blind tests, it was preferred over Sonnet 4.6, and it was trained from scratch without distillation from third-party models.
• MAI-Code-1-Flash: A 5-billion-parameter agentic coding model optimized for GitHub Copilot and VS Code.
• MAI-Image-2.5: Supports text-based image generation and editing, and includes an efficient Flash variant.
• MAI-Transcribe-1.5: A SOTA-level transcription model supporting 43 languages, 5 times faster than previous models.
• MAI-Voice-2: Supports natural voice generation for 15 languages, with the ability to adapt a voice from short samples.

Of particular note is the Frontier Tuning technology. This is a technique that uses reinforcement learning (RL) in real-world environments to make AI fully adapt to specific workflows.

Through RLE (Reinforcement Learning Environments), users can leverage an organization's actual internal work data and decision-making processes for model training. This allows a company's unique knowledge to be integrated into the model, simultaneously improving both performance and efficiency.