Mage-Flow: Microsoft's 4B Native-Resolution Image Model

Microsoft Asia has released Mage-Flow, a compact 4B-parameter generative stack for efficient text-to-image generation and instruction-based image editing. Instead of scaling to tens of billions of parameters, Mage-Flow reaches state-of-the-art-competitive quality through careful tokenizer-backbone-system co-design, keeping it fast, memory-light, and easy to fine-tune under realistic compute budgets.

• Model: microsoft/Mage-Flow on Hugging Face
• Paper: arXiv 2607.19064
• Code: github.com/microsoft/Mage
• License: MIT

What It Does

Mage-Flow is built from two shared, co-designed components:

• Mage-VAE — a lightweight, high-fidelity latent tokenizer (one-step diffusion encode/decode with anchor-latent KL regularization). Matches FLUX.2-VAE reconstruction fidelity while using ~12x / ~22x fewer encode / decode MACs per pixel.
• NR-MMDiT — a shared 4B Native-Resolution Multimodal Diffusion Transformer with rectified flow matching in the Mage-VAE latent space, using Qwen3-VL as the text encoder.

Together, these power two model instantiations — Mage-Flow for text-to-image and Mage-Flow-Edit for instruction-based image editing — each shipping in Base, RL-aligned, and 4-step Turbo variants.

Key Features

• Compact yet competitive — a single 4B family that matches or beats much larger open systems (Qwen-Image 20B, Z-Image 6B, FLUX.2 32B, FireRed-Image-Edit 20B)
• Native resolution — one checkpoint generates from 512 to 2048 on any aspect ratio, including extreme 4:1 (e.g. 512x2048, 2048x512)
• System-level speed — native-resolution packing + fused CUDA kernels raise MFU from ~33% to ~77% (~2.5x faster training); CFG's conditional/unconditional branches run in one packed forward
• Full model family — 6 checkpoints on Hugging Face covering generation and editing at Base, RL-aligned, and Turbo levels
• Interactive latency — at 1024x1024 on a single A100: Mage-Flow-Turbo 0.59 s/image, Mage-Flow-Edit-Turbo 1.02 s/edit, peak memory ~18-20 GB
• Versatile editing — supports semantic content editing, appearance transformation, image restoration, and structure-aware outputs within a unified pipeline

Model Zoo

All checkpoints are self-contained diffusers-style repos on Hugging Face:

Model	Task	Variant	Steps	Link
Mage-Flow-4B-Base	text-to-image	Base	30	🤗 Hugging Face
Mage-Flow-4B	text-to-image	RL-aligned	20	🤗 Hugging Face
Mage-Flow-4B-Turbo	text-to-image	Few-step distilled	4	🤗 Hugging Face
Mage-Flow-Edit-4B-Base	editing	Base	30	🤗 Hugging Face
Mage-Flow-Edit-4B	editing	RL-aligned	30	🤗 Hugging Face
Mage-Flow-Edit-4B-Turbo	editing	Few-step distilled	4	🤗 Hugging Face

Performance Highlights

Mage-Flow achieves GenEval 0.90 — the highest among all open-source models, surpassing FLUX.2 (0.87), Qwen-Image (0.87), and Z-Image (0.84). On DPG-Bench, Mage-Flow-Base scores 86.26, competitive with much larger models.

For image editing, Mage-Flow-Edit-Turbo scores GEdit-EN 8.271 and GEdit-CN 8.264, ranking first or second among all open-source editing models across multiple benchmarks.

Availability

All Mage-Flow checkpoints are released under the MIT license on Hugging Face, making them freely available for research and commercial use. At the time of writing, there is no native ComfyUI support for Mage-Flow yet — the model requires the official Python API, CLI, or Gradio app from the Mage repository.

See the project page for more details, galleries, and benchmarks.