Microsoft Research · 2026-07-22 · major
Microsoft Mage-Flow — 4B native-resolution image model that keeps up with 20B systems
Microsoft Research released Mage-Flow, a 4B open-weight text-to-image and instruction-editing model built on a lightweight tokenizer plus a rectified-flow diffusion transformer that generates 1024x1024 in 0.59 seconds on one A100.
A 4B image model from Microsoft Research that generates or edits at any resolution and matches open systems five times its size.
Key specs
| Parameters | 4B |
|---|---|
| 1024^2 latency | 0.59s (A100) |
| Gen eval | 0.9 |
Quick facts
| Maker | Microsoft Research |
|---|---|
| Parameters | 4B (Base, RL-aligned, Turbo, Edit variants) |
| License | MIT |
| Resolutions | 512 to 2048 px, arbitrary aspect ratios up to 4:1 |
| Latency | 1024x1024 in 0.59 s on A100 (Turbo) |
| Availability | Weights on Hugging Face, code on GitHub, paper on arXiv |
Benchmarks
What is it?
Microsoft Mage-Flow is a compact, open-weight foundation model for text-to-image generation and instruction-based image editing. The family includes a Base variant, an RL-aligned checkpoint, a 4-step Turbo distillation, and dedicated editing weights — all 4B parameters and released on Hugging Face under an MIT license.
How does it work?
Two co-designed pieces do the heavy lifting inside Mage-Flow. Mage-VAE is a lightweight latent tokenizer that improves reconstruction efficiency by an order of magnitude; the Native-Resolution Multimodal Diffusion Transformer sits on top and is trained with rectified flow matching so it can accept any input resolution up to 2048 pixels without retraining.
Why does it matter?
Mage-Flow reports a GenEval score of 0.90 and generates 1024x1024 images in 0.59 seconds on a single A100 — enough to match or beat Qwen-Image 20B, Z-Image 6B, and FLUX.2 32B at a fraction of the parameter count. Mage-Flow makes state-of-the-art open image generation practical on a single accelerator budget for research labs and hobbyists.
Who is it for?
image-model researchers, indie generative-AI teams, and open-source diffusion tooling maintainers
Frequently asked questions
- What tasks can Microsoft Mage-Flow do?
- Mage-Flow is one 4B family that handles both text-to-image generation and instruction-based image editing. Mage-Flow accepts natural-language edit commands (change, restore, insert), works on multiple input images, and generates natively at any resolution from 512 to 2048 pixels, including extreme aspect ratios like 4:1.
- How does Mage-Flow beat larger image models with only 4B parameters?
- Mage-Flow pairs a lightweight, high-fidelity latent tokenizer called Mage-VAE with a native-resolution multimodal diffusion transformer trained with rectified flow matching. Co-designing the tokenizer and backbone lets Mage-Flow match or beat Qwen-Image 20B, Z-Image 6B, FLUX.2 32B, and FireRed-Image-Edit 20B at a fraction of their size.
- Is Mage-Flow open source?
- Yes. Microsoft Research shipped Mage-Flow under an MIT license on July 22, 2026, with weights for the Base, RL-aligned, Turbo, and Edit variants on Hugging Face, code on github.com/microsoft/Mage, and the paper on arXiv (2607.19064). SafeTensors weights are compatible with the Hugging Face Diffusers library out of the box.
- How fast is Mage-Flow?
- Mage-Flow-Turbo generates a 1024x1024 image in 0.59 seconds on a single A100 GPU. Mage-Flow-Edit-Turbo runs an instruction edit on a 1024x1024 image in 1.02 seconds. The paper also reports a 2.5x training-throughput improvement over comparable systems thanks to the Mage-VAE tokenizer.
Try it
huggingface.co/microsoft/Mage-Flow