Microsoft Asia Open-Sources 4B Image Model Mage-Flow

Microsoft Asia released the new image generation and editing model family, Mage-Flow, on Hugging Face. The series includes three versions—Mage-Flow, Mage-Flow-Turbo, and Mage-Flow-Edit—with a core parameter count of 4 billion (4B) and is open-sourced for free under the MIT license. It quickly trended on Hugging Face, drawing significant community attention.

Confirmed

Mage-Flow is a compact model stack based on the DiT architecture. Instead of merely chasing benchmark scores, its focus is on streamlining training, fine-tuning, deployment, and inference efficiency. It utilizes the lightweight Mage-VAE latent space tokenizer and supports native resolution image generation and editing from 512×512 up to 4K. In terms of inference speed, the Mage-Flow-Turbo variant performs exceptionally well, requiring only 4 steps and less than 1 second to complete image editing at a 1024×1024 resolution, achieving interactive-level response times.

Why it matters

According to various community authors, Mage-Flow's core positioning is a "small model with quality close to large models." It supports not only basic text-to-image generation but also instruction-based image editing. This open-source strategy, balancing high quality with high efficiency, makes it a highly notable option in the lightweight image generation space.

2026-07-21 ~ 2026-07-23 · 12 related posts

Primary sources

6 near-duplicate retellings: multimodalart · multimodalart · multimodalart · multimodalart · multimodalart · pmttyji