MAI-Image-1

What it is

MAI-Image-1 is the first image generation model that Microsoft built entirely in-house. Microsoft AI announced it on 13 October 2025.^[1][2] A model of this kind takes a written prompt and returns a matching picture: during training it sees large collections of images paired with descriptions, and at generation time it starts from noise and gradually shapes that noise into an image that fits the prompt. "In-house" here means the model was developed by Microsoft's own team rather than supplied by an outside party. On LMArena, a public leaderboard where image models are compared through user votes, it debuted among the top ten text-to-image models.^[1]

What Microsoft says it does well

Microsoft described the model as strong at photorealistic imagery, singling out lighting effects such as bounce light and reflections, and also landscapes.^[1] Bounce light is light that has already struck one surface before reaching the subject; together with reflections it is one of the cues viewers use, often without thinking about it, to judge whether a rendered scene looks physically plausible.

Microsoft also presented the model as a balance between speed and quality, saying that users can iterate faster than with many larger, slower models.^[1] The company said it drew on feedback from creative professionals during development and worked to avoid outputs that look repetitive or fall into a generically stylised look.^[1]

Where it was available

From 4 November 2025 Microsoft rolled the model into Bing Image Creator, where it sat alongside DALL·E 3 and GPT-4o, and into Audio Expressions Story Mode in Copilot.^[1] From the same date it became available in every country with access to Bing Image Creator and Copilot Labs.^[1] Microsoft described the intended workflow in practical terms: get ideas onto the screen faster, iterate quickly, then move the work into other tools to keep refining it.^[1]

Where it sits in the line

Microsoft framed MAI-Image-1 as the next step after its first two in-house models, which arrived in August 2025, and as groundwork for more creative experiences inside its products.^[1] It did not stay at the frontier for long: it was succeeded by MAI-Image-2 in April 2026 and by MAI-Image-2.5 in June 2026.^[2]