# Microsoft's Mage-Flow packs image editing into 4B

> A compact model handles both text-to-image generation and instruction-based edits at native resolution, under a permissive MIT license.

Published by The Open Weights on Jul 21, 2026. Canonical: https://theopenweights.com/news/mage-flow-c5wr

## Key facts

- Company: Microsoft
- Model: Mage-Flow
- Version: 1.0
- Category: Text → Image
- Modalities: Text → Image, Image Editing
- License: MIT (Open weights, commercial use allowed)
- Parameters: 4B
- Architecture: MageFlowPipeline
- Precision: BF16
- Weights size: 8.2 GB
- Significance: notable
- Published: 2026-07-21
- Last verified: 2026-07-28
- Hugging Face: https://huggingface.co/microsoft/Mage-Flow
- Canonical URL: https://theopenweights.com/news/mage-flow-c5wr

Microsoft has released [Mage-Flow](https://huggingface.co/microsoft/Mage-Flow), a compact 4-billion-parameter model that handles both text-to-image generation and instruction-based image editing. Notably, it operates at native resolution and ships under a permissive MIT license.

The pairing of generation and editing in a single model is the headline feature. Rather than treating image creation and modification as separate tasks requiring separate tools, Mage-Flow aims to do both, following natural-language instructions to adjust an existing image or produce one from scratch.

## Why it matters

Size is the story here. At 4B parameters, Mage-Flow is small enough to run on more modest hardware than the sprawling diffusion systems that have dominated the space, which matters for developers who want local, self-hosted image tools without cloud dependencies.

- Dual capability: text-to-image plus instruction-based editing
- Native-resolution output
- MIT license, allowing commercial use and modification

The release is a first version for the family, with technical details outlined in the accompanying [paper](https://huggingface.co/papers/2607.19064). Whether Mage-Flow's compact footprint translates into competitive quality against larger open models will become clearer as the community puts it to work.

## Get the model

- [Hugging Face](https://huggingface.co/microsoft/Mage-Flow)

## Sources

- [microsoft/Mage-Flow](https://huggingface.co/microsoft/Mage-Flow) — Hugging Face, Jul 21, 2026

---
Source: The Open Weights (https://theopenweights.com/). Aggregated and written by Claude, curated by humans. Cite as: "Microsoft's Mage-Flow packs image editing into 4B", The Open Weights, Jul 21, 2026, https://theopenweights.com/news/mage-flow-c5wr