[Digital Today reporter Jinju Hong (홍진주)] Microsoft (MA) unveiled its in-house image-generation artificial intelligence (AI) model, MAI-Image-2.5.

On May 27 local time, online media outlet Gigazine reported that the model was ranked third in the world in the text-to-image generation category of the external evaluation service Arena.

The model is the latest version of Microsoft’s in-house AI series, MAI. After releasing its first in-house image-generation AI, MAI-Image-1, in October 2025, Microsoft unveiled an improved version, MAI-Image-2, and a speed-focused model, MAI-Image-2-Efficient, in April 2026. The newly released MAI-Image-2.5 highlights big gains in image quality and text-drawing performance compared with the immediately preceding MAI-Image-2.

Microsoft said MAI-Image-2.5 was launched at No. 3 on Arena, putting the external ranking at the forefront. Arena is a service in which humans evaluate output quality with AI model names hidden, and MAI-Image-2.5 took third place in the text-to-image generation category.

The focus of the performance upgrade is image completeness and instruction-following. Microsoft introduced the new model as performing inference by advancing subjects and scene structure, lighting, scale and spatial relationships. It also said it strengthened text rendering to accurately place user-specified phrases within an image. The emphasis is not only on image quality, but on precisely matching compositions to text instructions.

The comparison target was also clearly presented. MAI-Image-2.5 was evaluated as outperforming Google’s Nano Banana Pro in Arena’s assessment. It is a point where Microsoft presented external rankings and a competing model, not only internal comparisons, in explaining the new model’s competitiveness.

The product rollout schedule was also disclosed. Microsoft plans to make MAI-Image-2.5 available within 2 weeks in its generative AI service, MAI Playground. The new model is expected to move beyond the research stage and link directly to an actual service soon.

This trend shows Microsoft is rapidly segmenting its lineup of in-house image-generation models. After raising baseline performance with MAI-Image-2 and supplementing speed with MAI-Image-2-Efficient, it has now added MAI-Image-2.5 with enhanced image quality and text expression. By also signaling service deployment, it is moving to expand both verification of model competitiveness and user touchpoints at the same time.