Check where you can use the weights before you set aside disk space for them.
MiniMax published downloadable H3-Image weights for image generation and editing. The model card lists the release on November 10, and the company announcement is dated November 11. Anyone evaluating the download must first decide whether the license covers the intended use and whether the infrastructure can store and run the model.
Budget for the encoder as well as the checkpoint
MiniMax lists 68.65 GB for the main files and 66.7 GB for the separate text encoder. A fresh setup that needs both starts at about 135.35 GB. The optional 17.9 GB prompt enhancer brings that to about 153.25 GB. An existing cache can change how much you need to download. Allow room for the cache and intermediate copies if your installation keeps them on the same volume.
The figures describe files on disk. Pipeline runtime also depends on GPU memory, precision, image dimensions, and the implementation. MiniMax reports that its own setup generated a 4,096 × 4,096 image in under 10 seconds. Use that as a vendor measurement and time your own run on your intended hardware.
Check the terms for your deployment
The H3-Image Community License excludes the United States, European Union, United Kingdom, and South Korea from its permitted territory. It also contains provisions for hosted services. Read the current agreement before downloading weights for a team or planning a service that will expose the model to customers. MiniMax lists a separate hosted app, and its model card links to API documentation.
For a first evaluation, save the same input and edit prompt for two runs. Record the model revision, output dimensions, elapsed time, and which parts of the original image changed. If the work involves repeated edits, reuse the same image across runs. Our Seedream 5.5 coverage describes a similar way to spot accumulated changes.
Abdessalam Alaoui
I'm Abdessalam Alaoui, founder of Musthave.AI. I write practical guides to AI models, tools, generative media.
