Confirm where you are before downloading

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

These weights are licensed by MiniMax under the MiniMax-H3-Image Community License, which covers use and distribution only outside the European Union, the United Kingdom, the Republic of Korea and the United States of America. Read the LICENSE file in this repository before you download.

Log in or Sign Up to review the conditions and access this model content.

MiniMax-H3-Image for Mac (MLX, 8-bit)

Run it in one click with Phosphene: a free, open-source Mac app for offline AI pictures, video, speech and songs.

Will it run on your Mac? See which models run on your memory size and how long a clip takes: phosphene.video/mac-requirements.

Download Phosphene at phosphene.video (Mac app, or one-click install through Pinokio), open Images and pick MiniMax-H3-Image.

This is MiniMax-H3-Image converted to run natively on Apple Silicon with MLX: the image transformer in 8-bit (group size 64), with its per-block timestep projections precomputed for the model's six steps. About 22 GB instead of 66 GB. Everything runs on your Mac; nothing is sent anywhere.

Benchmark (M4 Max, 64 GB)

Whole picture, wall clock, including model load:

Size Time Peak memory
4 MP about 2 min (119 s) 27.5 GB
8 MP about 4 min (227 s) 27.5 GB
16 MP about 8.5 min (513 s) 27.5 GB

6 steps, no guidance. Needs a Mac with 48 GB of memory or more. Strong at hands, skin and exact text on posters. Longer, detailed English prompts work best.

Edit and combine pictures

Since Phosphene 5.0.19 the same weights also edit your own pictures:

  • Edit one picture: give it a photo and say what to change ("make it night", "put her in a red coat"). The rest of the picture stays.
  • Combine pictures: give it two or more pictures (up to 16) and describe the result, for example a person from one photo in the place from another.
  • Match the reference: keep the layout and framing of your picture while restyling it.

Measured on the same Mac (M4 Max, 64 GB), 4 MP, whole picture including model load:

Job Time Peak memory
Text only about 2 min (119 s) 27.5 GB
Edit one picture about 5 min (277 to 291 s) 29.3 GB
Combine two pictures about 8 min (493 s) 29.3 GB

In Phosphene: open Images, pick MiniMax-H3-Image, add your pictures, write what you want and press Generate.

Files

  • dit/: the image transformer, 8-bit MLX, 12 shards, sha256 per file in dit/ASSET_SHA256.json
  • vae/: the image decoder
  • pdd_fused_heads.safetensors, pdd_config.json: the per-step output heads and schedule from the original release
  • LICENSE, NOTICE: the original licence and notice

The text encoder is the one from MiniMax-H3 (Qwen3-VL-32B); Phosphene's H3 install already includes it.

Licence

MiniMax-H3-Image is licensed under the MiniMax-H3-Image Community License Agreement, Copyright (c) 2026 MiniMax. All Rights Reserved. The full licence is in LICENSE, and it applies to these files and to anything you make with them, including the Acceptable Use Policy in its Exhibit A. The licence does not cover the European Union, the United Kingdom, the Republic of Korea or the United States of America.

The transformer and decoder files here were modified from the original (8-bit quantisation, precomputed timestep projections); each modified file says so in its metadata.

Powered by MiniMax-H3-Image (MiniMax). Pictures made with this model are AI-generated; please label them when you publish them.

Report misuse: open a discussion on this repository.

Downloads last month

-

Downloads are not tracked for this model. How to track
MLX
Hardware compatibility
Log In to add your hardware

8-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for Mrbizarro/MiniMax-H3-Image-MLX-8bit

Quantized
(32)
this model