Microsoft open-sources Mage-Flow, a 4B image generation and editing model under MIT license

Microsoft released Mage-Flow, a 4-billion-parameter open-source model for text-to-image generation and instruction-based image editing, under a permissive MIT license that lets developers use it commercially without restriction.

The pitch is capability at a small size. Mage-Flow generates and edits images at native resolutions from 512 to 2,048 pixels, and Microsoft reported a GenEval score of 0.90 on text-to-image generation, which it said is the highest among the open-source models it tested. On the ImgEdit-Bench editing benchmark it scored 4.34.

Architecturally, the system pairs a lightweight latent tokenizer Microsoft calls Mage-VAE with a Native-Resolution Multimodal Diffusion Transformer, trained using a technique known as rectified flow matching. It ships in three variants: Base, a reinforcement-learning-tuned RL version, and a 4-step Turbo build for faster generation.

Microsoft claims the 4-billion-parameter model holds its own against far larger systems, including a 32-billion-parameter Flux.2 variant and a 20-billion-parameter Qwen-Image variant. Those benchmark results are Microsoft’s own and have not been independently verified.

Microsoft released the weights under the MIT license, with code on GitHub and an accompanying research paper on arXiv.