How to run MageFlow locally.

Which file.

Mage-Flow is the image half of Microsoft’s Mage research family, released on 22 July 2026: a 4B MMDiT with its own one-step VAE, prompted through Qwen3-VL 4B. There are six checkpoints, three for making images and three for editing them.

• Mage-Flow (mage_flow_bf16) is the RL-aligned model and the one to start with. 20 steps at CFG 5.
• Mage-Flow Turbo makes an image in exactly 4 steps at CFG 1, no negative prompt.
• Mage-Flow Base is the plain pretrained model, for training and variety. 30 steps.
• Edit, Edit Base and Edit Turbo take a reference picture and an instruction.

The licence is MIT, which allows commercial use. Microsoft’s Responsible AI note adds that the models are released for research and not meant for products or services. That note is guidance, not a licence term, but read it before you build on it.

Files you need.

One model, one encoder, one VAE. Comfy-Org ships each model in bf16 and in int8_convrot, a newer int8 format that is half the size.

• Model mage_flow_int8_convrot.safetensors ComfyUI/models/diffusion_models/ or mage_flow_bf16.safetensors, 8.2 GB
  4.2 GB Download
• Turbo mage_flow_turbo_int8_convrot.safetensors ComfyUI/models/diffusion_models/ optional. bf16 is 8.2 GB
  4.2 GB Download
• Text encoder qwen3vl_4b_bf16.safetensors ComfyUI/models/text_encoders/
  8.9 GB Download
• VAE mage_flow_vae_bf16.safetensors ComfyUI/models/vae/
  0.3 GB Download

The VAE is Mage-Flow’s own one-step decoder, and the templates load it by name. No GGUF builds exist.

models/
├── diffusion_models/
│   └── mage_flow_int8_convrot.safetensors
├── text_encoders/
│   └── qwen3vl_4b_bf16.safetensors
└── vae/
    └── mage_flow_vae_bf16.safetensors

What fits your computer.

At 1024 × 1024. The encoder runs first and ComfyUI moves it out of the way before sampling, so the model file sets the limit. Few people have posted numbers yet, so treat the lower tiers as estimates from file sizes.

• 8 GBTight

  int8 model (4.2 GB). The 8.9 GB encoder is bigger than the card, so ComfyUI keeps part of it in system RAM while it encodes.

• 12 GBFits

  int8 comfortably. bf16 fits with little room at larger sizes.

• 16 GBFits

  bf16. An underclocked RTX 5070 Ti runs Turbo quickly.

• 24 GBFits

  bf16 at 1536 to 2048 px, where Mage-Flow looks best. Microsoft measured an 18 to 20 GB peak in its own code.

• Older GPUsUpdate

  Cards without bf16, such as a GTX 1660, gave corrupt images until ComfyUI PR #15081. Update ComfyUI.

• MacUntested

  No Mac reports yet. The bf16 files are the safer bet on Apple Silicon; int8 files have caused trouble on MPS with other models.

• 8 GB graphics cards
• 12 GB cards
• 16 GB cards
• Mac with Apple Silicon

Set it up.

1. 1

   Update ComfyUI

   Mage-Flow is native from ComfyUI 0.29.0. Older versions don’t have TextEncodeMageFlowEdit. ComfyUI Desktop updates itself; in a manual install:

   Terminal, in the ComfyUI folder Copy

   git pull
   pip install -r requirements.txt

2. 2

   Download the three files

   Model, encoder and VAE from Comfy-Org/Mage-Flow, using the list above.

3. 3

   Put them in their folders

   Model in diffusion_models, encoder in text_encoders, VAE in vae. Restart ComfyUI.

4. 4

   Open a template

   Pick Mage-Flow: Text to Image or Mage-Flow Turbo: Text to Image in the template browser. The Mage-Flow-Edit templates are for the edit files.

5. 5

   Check the loaders

   Load CLIP gets qwen3vl_4b_bf16 with type mage. TextEncodeMageFlowEdit encodes the prompt and makes the latent in one node, so there is no separate empty-latent node.

6. 6

   Set steps and run

   The text-to-image template starts at 30 steps. That suits Base; for the regular Mage-Flow file, 20 is Microsoft’s setting. Turbo stays at 4 steps and CFG 1.

Settings that work.

Mage-Flow uses the plain KSampler. ComfyUI sets its shift to 6 automatically, which matches Microsoft’s, so there is no ModelSampling node to add.

Mage-Flow and Base

Steps

20Base: 30

CFG

5

Sampler

euler

Scheduler

simple

Shift

6, built in

Size

1024 to 2048multiples of 16

Negative

yes

Text encoder

Load CLIP, mage

Turbo

Steps

4

CFG

1

Sampler

euler

Scheduler

simple

Any aspect ratio works, up to 4:1. Comfy’s template note suggests 1024 for tests and 1536 to 2048 for quality, and users report much better results at 2048 × 2048 than at 1024. Turbo is strict: Comfy’s template note says it must run at 4 steps and CFG 1.

How fast.

Microsoft’s figures, in its own code rather than ComfyUI.

When it goes wrong.

Missing node TextEncodeMageFlowEdit

ComfyUI is older than 0.29.0. Update it, or switch to the nightly build.

Turbo gives a cloudy mess after the first step

A bug on GPUs without bf16 (GTX 16xx and similar), fixed in ComfyUI PR #15081. Update ComfyUI, and keep Turbo on euler with simple.

Turbo images look soft or muddy

Turbo needs exactly 4 steps and CFG 1. Also try 1536 px or more.

Edit ignores the reference image

Reported on AMD with ROCm, from the Comfy Kitchen attention backend. Switch ComfyUI to another attention backend.

The microsoft/Mage-Flow download link returns 401

Microsoft removed those repos. Download from Comfy-Org/Mage-Flow instead.

Questions.

Sources: Mage-Flow README [1], Microsoft Mage repository, Comfy-Org Mage-Flow files, Turbo and resolution thread, TextEncodeMageFlowEdit thread, ROCm reference thread, stable-diffusion.cpp note on the removal.