whatmodel.sh / image / Mage-Flow

Mage-Flow

+ contribute
Update model informationAdd a place to run itAdd a test resultPropose a benchmarkCorrect a benchmark scoreEdit data on GitHub

Mage-Flow is an open-weights image model from Microsoft, released July 2026. Its published size is 4B parameters.

Model information sources

Release date ↗ · Parameter count ↗

Maker's paper (Microsoft, arXiv 2607.19064). 4B = the NR-MMDiT backbone shared by Base, RL-aligned and Turbo text-to-image variants (README names them Mage-Flow-4B-*); Mage-VAE and any text encoder are not included.

69.2±19.7#1 of 3 image modelsfrom 9 benchmark results across 1 source

Score within the CVTG-2K word accuracy comparison set only.

$ how it compares

Scored image models released closest to Mage-Flow, out of 197 with a score, then more from Microsoft.

#modelscorewhereRAMreleased
4MAI-Image-2.6 Microsoft AI81±2comparison only2026-08
7Reve 2.1 Reve78±2api $0.25/image2026-07
8Muse Image Meta76±2api $0.010/image2026-07
11Qwen-Image-3.0-Pro Alibaba75±2api $0.040/image2026-07
12Seedream 5.0 Pro ByteDance Seed75±2api $0.050/image2026-07
more from Microsoft
110VQ-Diffusion (ITHQ) Microsoft48±15local~4.1GB2022-11

see all image models · find the ones that fit your machine

$ what it's good at

text to image█····█·····█ #1 of 3 · 76% of its score · 9 tests
testresultvs the others in its rundifficultycountswhy
GenEval lab90.0% ✎▇▃▅▇▁▄▅▃█▃▃▅ beats 39/39very easy · avg 87%▇▇▇▇▇▇▇▇▇ 22%sharp · maxed out · small run
CVTG-2K word accuracy lab88.7% ✎▄▄▄█▄▄·▄···▄ beats 7/8very easy · avg 87%▇▇▇▇ 10%small run
TIIF-Bench testmini (short prompts) lab82.2% ✎█·█········█ beats 0/2easy · avg 83%▇▇▇ 7.8%small run
TIIF-Bench testmini (long prompts) lab84.7% ✎█······█···█ beats 2/2easy · avg 84%▇▇▇ 7.6%small run
LongText-Bench (CN) lab82.3% ✎▃·······████ beats 3/12easy · avg 81%▇▇▇ 7.1%small run
LongText-Bench (EN) lab94.4% ✎▃·····▅██▅▃▃ beats 8/12very easy · avg 92%▇▇ 6.1%maxed out · small run
DPG-Bench lab86.5% ✎▄·▂·▂▆█▄▂▄▂▂ beats 11/17very easy · avg 86%▇▇ 5.3%small run
OneIG-Bench (CN) lab50.5% ✎▃·▃█▃·▃··▃·▃ beats 1/8medium · avg 50%▇▇ 5.2%at its level · small run
1 more test
testresultvs the others in its rundifficultycountswhy
OneIG-Bench (EN) lab53.6% ✎▃······▃▅█▅▃ beats 5/9medium · avg 53%▇▇ 5.1%at its level · small run

$ where to run it

Other ways to run it

version

  1. Run it as published · needs ~13.8 GB · from github.com/leejet
    sd-cli --diffusion-model models/mage_flow_bf16.safetensors --llm models/Qwen3VL-4B-Instruct-Q4_K_M.gguf --vae models/mage_flow_vae_bf16.safetensors -p "a lovely cat holding a sign says 'mage.cpp'" --cfg-scale 5.0 --steps 20 --diffusion-fa -v --offload-to-cpu

Official weights and documentation: github.com/microsoft ↗

$ how the score was worked out · Mage-Flow: An Efficient Nativ… 76%

Each result implies a score on its own, after adjusting for how hard its test is and who ran it. The score is their weighted average, with a small pull toward the middle for thinly tested models. Each test's weight is its counts column above; here it's added up by who ran the tests. How it works

sourceresultsshare of the scorelicence
Mage-Flow: An Efficient Native-Resolution Foundation Model for Image Generation and Editing · lab-reported9▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇ 76%CC-BY-4.0
prior · pull toward the middle▇▇▇▇▇▇▇ 24%
weighted average= 69.2

$ improve this page

Update model informationAdd a place to run itAdd a test resultPropose a benchmarkCorrect a benchmark scoreEdit data on GitHub

Use a short GitHub form or edit the data through a pull request. Every number needs a source; here's which ones we can use.

data updated · sources and licences · json