MiniMax-H3-Image on Mac: 135GB in BF16, Not Licensed in the US

MiniMax published MiniMax-H3-Image on 10 November 2026: one open-weights checkpoint that generates images from text and edits up to 16 reference images, at about 16 megapixels natively. The official weights are 68.65 GB in BF16, and they need the 66.7 GB text encoder of MiniMax-H3 on top, so the full-precision set is about 135 GB. Before the hardware question there is a legal one: the licence does not cover the United States, the European Union, the United Kingdom or South Korea, so readers there have no licence to run it at all, locally or not.

TL;DR: In BF16 the model needs about 135 GB (DiT and VAE 68.65 GB, text encoder 66.7 GB), which a 512GB Mac Studio holds with room to spare. Smaller builds exist: Comfy-Org's INT8 diffusion model is 20.979 GB and a community Q4_K_M GGUF is 11.430 GB. On a Mac, the routes on 11 November are stable-diffusion.cpp, which merged support on 10 November, and a preview in the Phosphene app that asks for 48 GB of memory. We found no published Mac timing yet. The licence excludes the US, the EU, the UK and South Korea, needs written permission above 1 million US dollars of annual revenue or for any hosted service, and forbids using outputs to improve another AI model.

Can you use MiniMax-H3-Image where you live?

Read this part first if you are in the US. The MiniMax-H3-Image Community License Agreement, dated November 10, 2026, defines the "Applicable Territory" as "worldwide, excluding the Excluded Territories", and the Excluded Territories are "the European Union, the United Kingdom, the Republic of Korea and the United States of America". The grant itself opens with "Solely within the Applicable Territory". The use restrictions add that you may not use the works "or any of their Outputs or results outside the Applicable Territory".

That covers running the model on your own Mac, not only hosting it. The licence offers a way in for those countries: people there are "welcome to contact us about obtaining a license". Until MiniMax grants one, the download is public but the use is not licensed in those four places.

Three more terms matter for everyone else:

The revenue line is much lower than the one on the video model. Our MiniMax H3 on Mac guide covers the video licence, whose threshold is 20 million US dollars. For the image model it is 1 million, and it counts annualized revenue of you and your affiliates, not only the product built on the model. The optional prompt expander is covered too: the licence says it "is part of MiniMax-H3-Image and is accordingly subject to this Agreement", even though it is derived from Qwen3.8-27B under Apache 2.0.

What MiniMax shipped

The model card describes a diffusion transformer, or DiT, the network that turns noise into the image step by step. Here it has the H3 architecture without the audio layers, and the model uses the MiniMax-H3 text encoder unchanged.

The 6 steps are fixed. In the official request schema, num_inference_steps must be omitted or set to 6, guidance_scale must be omitted or set to 1, and other values return HTTP 400. A negative prompt is rejected. The card also says the model expects "a detailed English description of the whole image", which is why MiniMax ships a prompt expander.

How big are the files?

Sizes below come from the model card and from the Hugging Face API on 11 November 2026, in decimal gigabytes.

One trap: on 11 November the repository holds more than the card's table. The Hugging Face API lists a diffusers-format copy of the DiT (66.289 GB) and of the VAE (2.337 GB), plus a text_encoder/ folder of 66.727 GB, for 204.02 GB in total. model_index.json still loads the encoder from MiniMaxAI/MiniMax-H3. So a plain download of the whole repository pulls about three times what one pipeline needs. Pick one format and use the card's own commands, which download the repository and then only the text_encoder/ folder of MiniMax-H3.

If you already run the MiniMax H3 video model, the card says the encoder "is not downloaded again".

The distance to a 512GB Mac Studio

This model fits, by a wide margin. The BF16 set is 68.65 GB for the model plus 66.7 GB for the text encoder, so 135.35 GB of weights. Against Apple's 512GB, counted here as 512 GB in decimal units, that is about 26 percent of the machine. Adding the optional prompt expander brings it to 153.25 GB, about 30 percent. These are weights alone, before activations at 16 MP and before macOS. Apple's 512GB is binary, so the real headroom is slightly larger than this calculation shows.

The question on a Mac is not whether 135 GB fits a 512GB machine. It is which software runs the model on Apple Silicon, and what the smaller machines can load.

Which builds are smaller?

The heavy part shrinks a lot. Comfy-Org, the team behind ComfyUI, published a pruned DiT and quantized encoders on day 0, and community GGUF files followed within hours. File sizes come from the Hugging Face API on 11 November 2026.

The ComfyUI tutorial loads four files: the INT8 DiT, the INT8 text encoder, the 9B prompt expander and the FP16 VAE. Its text-to-image template opens at 4 MP "so low-VRAM machines do not run out of memory". It needs ComfyUI 0.45.0 or later and template pack 0.11.104 or later.

Our calculation for that set: 20.979 + 27.141 + 7.344 + 1.169 = 56.63 GB of weights. The text-to-image template ships with the expander switched off, which leaves out its 7.344 GB. On Apple Silicon, ComfyUI runs on PyTorch MPS rather than MLX. The W4A8 and NVFP4 files are built for NVIDIA kernels, and we found no source showing that they run on MPS.

How do you run it on a Mac?

MiniMax recommends SGLang, vLLM-Omni, diffusers and ComfyUI. The vLLM-Omni recipe lists NVIDIA and AMD accelerators such as the H200, the RTX PRO 6000 Blackwell and the Instinct MI350X. None of the four names a Mac. Three routes reach Apple Silicon on 11 November:

• stable-diffusion.cpp. PR #2324, "feat: add MiniMax-H3-Image support", was merged on 10 November 2026. The Hearthscale card says its GGUF files "load in stable-diffusion.cpp from master-1088, the first build that reads MiniMax H3-Image, with no patch". stable-diffusion.cpp has a Metal backend, so this is the merged upstream route for a GGUF on a Mac. The timings on that card come from an NVIDIA RTX 3080, not from a Mac.
• Phosphene. The free Mac app added MiniMax-H3-Image in version 5.0.15 on 10 November, as a preview. It does text to image only, at 4, 8 or 16 megapixels, and it "needs a Mac with 48 GB of memory or more". The first use downloads about 69 GB. Its engine notes in the same commit say each DiT shard is quantized to Q8 on your Mac, giving a pack of about 21 GB. The text encoder, the DiT and the decoder run one after the other, "so a 64 GB Mac never holds all three". Its time estimates are still marked as placeholders awaiting a benchmark.
• ComfyUI with the day-0 templates above, on PyTorch MPS.

An MLX conversion in BF16, VisualInference/MiniMax-H3-Image-MLX-bf16, appeared on 11 November. Its card is gated, so we could not read what software it targets. Treat it as a file listing, not a route.

We found no published Mac timing for MiniMax-H3-Image on 11 November 2026. When one is published with the machine, the settings and a link, it can go on a MiniMax-H3-Image page next to our image model pages, the way the MiniMax H3 video page cites published timings.

How much does the API cost instead?

For readers who would rather not download about 135 GB, MiniMax sells the model by the image. Crypto Briefing reports about $0.022 for a 1K image and about $0.069 for a 4K image, with promotional discounts until November 17, 2026. The API is a hosted service run by MiniMax, but the licence's territory clause still decides what you may do with the model and its outputs.

Bottom line

MiniMax-H3-Image is a big model with a small path. In BF16 it is about 135 GB with its text encoder, so a 512GB Mac Studio holds it easily. Pruned and quantized files bring the DiT down to 21 GB in INT8 and 11.430 GB in a Q4_K_M GGUF. On a Mac, stable-diffusion.cpp has merged support and Phosphene ships a 48 GB preview, but nobody has published a Mac timing yet. The licence is the harder limit: no use in the US, the EU, the UK or South Korea, written permission above 1 million US dollars of revenue, and no training other models on its outputs.

To compare with image models that already have Mac pages, see FLUX.2 dev and Z-Image Turbo, or our guide to local image generation on a Mac. If you are sizing the biggest machine, our 512GB Mac Studio guide covers what else fits.

FAQ

Can I use MiniMax-H3-Image in the United States?

Not under its public licence. The agreement excludes the United States of America, the European Union, the United Kingdom and the Republic of Korea, and it forbids using the model or its outputs outside the Applicable Territory. MiniMax invites people in those territories to contact it about a separate licence.

How much memory does MiniMax-H3-Image need?

In BF16, about 135 GB of weights: 68.65 GB for the DiT and VAE plus 66.7 GB for the text encoder from MiniMax-H3. Comfy-Org's INT8 diffusion model (20.979 GB) and INT8 text encoder (27.141 GB) cut that sharply, plus 1.169 GB for the FP16 VAE. Phosphene asks for a Mac with 48 GB or more and runs the parts one after the other.

Is the text encoder included in the download?

The model card says no: model_index.json loads it from the text_encoder/ folder of MiniMaxAI/MiniMax-H3, which is 66.7 GB. On 11 November the repository also carries its own text_encoder/ folder and a diffusers-format copy of the weights, 204.02 GB in total, so download selectively.

Does it run with llama.cpp or mlx-lm?

No. Those are language-model runtimes, and this is a diffusion model. The upstream route for GGUF files is stable-diffusion.cpp, which merged MiniMax-H3-Image support on 10 November 2026.

How many steps does it take?

Six. The checkpoint is distilled to 6 sampling steps with no classifier-free guidance, and the official server returns HTTP 400 for any other step count or guidance value.

Sources

• MiniMaxAI/MiniMax-H3-Image model card: 33.1B DiT, 50 layers, 6 steps without CFG, about 16 MP, 16 references, file sizes, text encoder not in the repository, recommended frameworks, request parameters.
• MiniMax-H3-Image Community License Agreement: November 10, 2026, excluded territories, the 1 million USD and hosted-service authorization, the ban on improving other models, the prompt expander note.
• Hugging Face API, MiniMaxAI/MiniMax-H3-Image with blob sizes: 204.02 GB in total on 11 November 2026, including the diffusers copy and the text_encoder/ folder.
• MiniMaxAI/MiniMax-H3-Image model_index.json: the text encoder loaded from MiniMaxAI/MiniMax-H3, subfolder text_encoder.
• Hugging Face API, MiniMaxAI/MiniMax-H3: the text_encoder/ folder, 66.7 GB per the MiniMax-H3-Image model card.
• MiniMaxAI/MiniMax-H3-Image-PE: the optional NVFP4 prompt expander, 27.36B parameters, 17.9 GB per the MiniMax-H3-Image model card.
• Comfy-Org/MiniMax-H3-Image and its blob sizes: pruned BF16, INT8 and W4A8 DiT files, the INT8 and NVFP4 text encoders, the 9B prompt expander, the FP16 VAE.
• ComfyUI MiniMax-H3-Image tutorial: ComfyUI 0.45.0, template pack 0.11.104, the four files, the 4 MP default canvas.
• Hearthscale/MiniMax-H3-Image-GGUF and its blob sizes: Q8_0 to Q3_K_M files, stable-diffusion.cpp master-1088.
• stable-diffusion.cpp PR #2324: MiniMax-H3-Image support, merged 10 November 2026.
• Phosphene 5.0.15 release notes and commit 2a608e3fd7: the 48 GB preview, the 69 GB download, the Q8 pack of about 21 GB, the staged run, the time estimates awaiting a benchmark.
• Crypto Briefing, 10 November 2026: API prices of about $0.022 (1K) and $0.069 (4K), the promotion until November 17.