生成式模型與開發者平台
Microsoft Brings MAI-Image-2.6 and Flash to Foundry, Unifying Image Generation and Multi-Reference Editing APIs
The MAI-Image-2.6 family uses a common set of Foundry interfaces for text-to-image generation, multi-reference image editing, web grounding, and output at up to 1.5K resolution. The Flash variant targets high-throughput workloads, although its claimed speed and efficiency advantages currently rely primarily on Microsoft’s own measurements.
On September 4, Microsoft added MAI-Image-2.6 and MAI-Image-2.6-Flash to Microsoft Foundry in public preview. Both models support text-to-image generation and image editing, with new capabilities including multiple reference-image inputs, the use of web content to augment generation context, dynamic aspect ratios, and output at up to 1.5K resolution. The standard model prioritizes accuracy, while Flash is designed for latency-sensitive and high-volume image-generation workflows.
For developers, the significant change is more than the arrival of another model competing on leaderboards. Foundry separates generation and editing into the `/mai/v1/images/generations` and `/mai/v1/images/edits` APIs, which support authentication through Entra ID or an API key. Global Standard is currently the available deployment type. The model version is listed as `2026-07-31`, but the service remains in Preview. The free tier includes no request quota, while the documented quota for the initial paid tier is only two requests per minute. Teams planning production batch workloads will therefore still need to request capacity and confirm regional availability.
Microsoft claims that Flash generates images 2.8 times faster than GPT-Image-2-Medium and delivers a 72% efficiency improvement. In terms of quality, MAI-Image-2.6 ranked second on the Arena text-to-image leaderboard on September 4, with an Elo score of 1332±7 based on 10,086 cumulative votes. This provides an independent signal based on blind user preferences, but it does not validate Microsoft’s latency or cost measurements, nor does it represent specific workloads such as Chinese text rendering, character consistency, or editing branded assets.
Organizations evaluating adoption should test the standard and Flash variants separately. They should hold prompts, the number of reference images, resolution, and safety settings constant while recording tail latency, retry rates, actual cost per image, and identity consistency after editing. Web grounding also introduces concerns around source freshness, copyright, and prompt injection, so generated outputs should not be treated as verified facts. Key issues to watch next include the production SLA, regional availability and data-processing boundaries, and whether Microsoft publishes reproducible speed-test conditions and more detailed safety evaluations.