
Microsoft AI’s MAI models explained for code, security, voice and image
A workload guide to Microsoft’s MAI models, covering Cyber, Code, Voice, Transcribe, Thinking and the Image 2.7 preview.
Which MAI model fits which job?
Microsoft’s model catalog separates software engineering, cybersecurity, speech, reasoning, and image tasks. MAI-Code-1.1-Flash is built into GitHub Copilot and VS Code. MAI-Cyber-1.1-Flash is used in Microsoft’s MDASH system for vulnerability discovery and remediation. Evaluate coding with representative repository tasks and passing tests. Evaluate security with reproducible findings, agreed scope and verified fixes. Microsoft AI model catalog
MAI-Transcribe-2 turns audio into text. MAI-Voice-2.1 produces speech. In a support workflow, score transcription against recordings with overlapping speakers, noise and the terms your customers use. For voice, measure how long the user waits for the first audio and how well the output handles names, numbers and longer responses. These are separate workloads even when they sit in the same product.


MAI-Thinking-1 is Microsoft’s reasoning model in the catalog. Start by testing a few realistic tasks with known answers and compare accuracy, latency, and cost per accepted result. For tool use, check the arguments and permission decisions in addition to the written answer.


Is MAI Image 2.7 ready for the same deployment as 2.6?
Microsoft announced MAI Image 2.7 in preview on October 27 with a focus on difficult image edits. It separately announced general availability for MAI Image 2.6 and 2.6 Flash in Microsoft Foundry. Check the availability of the specific version in your tenant and region. For a product imagery workflow, test whether logos, layout and existing design elements survive the edit, and have a reviewer score the result. Image 2.7 preview · Image 2.6 availability


How should businesses compare the models?
Use a small, fixed set of realistic examples with clear pass criteria. Record the model version, tool access, output quality, review time, and the cost of retries. Run the same examples on a baseline model before switching providers. Microsoft’s Foundry October roundup is a useful companion for checking where model and agent capabilities show up in the platform.
Take the next step
Bring one workflow and the systems it needs to access.
Let's Talk