MAI-Image-2.6-Flash

Microsoft's official MAI-Image-2.6 product page describes the model as a tool for creating and editing high-quality, design-ready images from text, images, web context, and visual references. The page emphasizes precise control over composition, style, and detail for both casual and professional users, with a try-in-Playground option. Multi-reference editing enables combining up to five visual references into a single coherent creative direction. The product page highlights features that the MAI-Image-2.6 family supports, including targeted object edits, reframing, layout adjustments, text updates, and artifact cleanup while preserving surrounding details. These capabilities underpin the family's design around how people create, with emphasis on consistency and control. The page serves as Microsoft's first-party entry point for the MAI-Image-2.6 line that includes the Flash variant.

MAI-Image-2.6-Flash launched in public preview on Microsoft Foundry with Arena rankings of No. 3 in image editing and No. 8 in text-to-image on the Artificial Analysis Image Arena. Pricing is set at $1.75 / $2.50 / $19 per million tokens for text-in / image-in / image-out respectively. Against the prior MAI-Image-2.5-Flash at the same price, the new model gains 69 Elo in text-to-image and 34 Elo in editing. The write-up details a shared 2.6 family feature set: up to five reference images per request, Bing Search grounding for current information, dynamic aspect ratios, and outputs up to 1.5K resolution. Both models use a faster diffusion stack with flow-matching training that can reduce sampling steps. Flash runs about 2.8x faster than GPT-Image-2-Medium with 72% greater efficiency, targeting high-volume workloads.

The official MAI-Image-2.6 / MAI-Image-2.6-Flash Model Card (dated September 4, 2026) documents a diffusion-based generative architecture using flow-matching loss for text-to-image synthesis and image-to-image editing. The model has 20B non-embedding parameters, a 32K-token context length, and supports outputs up to 2,359,296 pixels (1536x1536) with either dimension extendable. Training ran from April 17 to July 22, 2026, with the flagship released August 14 and Flash on September 4. The Model Card confirms MAI-Image-2.6-Flash is an optimized version of MAI-Image-2.6, accepting text and image inputs for editing workflows. Licensing follows product terms where deployed, such as Azure AI Foundry and MAI Playground. Developer Representative is Microsoft Ireland Operations Limited, with Microsoft Corporation as the named developer of the model.

Microsoft introduced MAI-Image-2.6-Flash on September 4, 2026, positioning it as a latency-optimized sibling to the flagship MAI-Image-2.6 for high-throughput production workloads. The Flash variant is reported to generate images 2.8x faster than GPT-Image-2-Medium with 72% greater efficiency, while supporting multi-image reference editing, web grounding, and dynamic aspect ratios. On benchmarks, MAI-Image-2.6 ranks No. 2 on Arena for both text-to-image and image editing. Both models became available to developers in Microsoft Foundry, giving users a choice between maximum precision (MAI-Image-2.6) and production speed (MAI-Image-2.6-Flash). The announcement describes shared family capabilities including reference-based editing, Bing-grounded web context, and flexible output dimensions. The release expands the MAI-Image line with a speed-focused tier aimed at cost-sensitive, latency-critical image generation and editing pipelines.