H3 Image

Run H3 Image Locally H3 Image Download and VRAM Guide

Every H3 Image Hugging Face version with its size, plus VRAM, tool and license notes.

MiniMax H3 Image is open-weight, so you can run H3 Image locally: download it from Hugging Face and load it in ComfyUI, Diffusers, SGLang or vLLM-Omni, or a community MLX build on a Mac. The official model is 68.65 GB in BF16 and needs a separate 66.7 GB text encoder; community GGUF and FP8 files are smaller. Below: each version's size, the settings the model expects and the license terms. We have not run these files ourselves.

  • Official download: 68.65 GB in BF16
  • H3 Image GGUF, FP8 and MLX versions
  • ComfyUI, Diffusers, SGLang, vLLM-Omni
  • License excludes US, EU, UK, S. Korea
0
A woman lying on a sheepskin rug in warm afternoon window light
Prompt

A casual phone photo of a woman in her mid-twenties lying back on a white sheepskin rug on a wooden floor beside a sunny window. Long dark hair spread out, one hand resting against her cheek, eyes half closed. She wears a short-sleeved floral summer dress and slouchy cream knit socks. Open paperback books and a few ripe apricots lie on the rug around her. Warm, hard late-afternoon sunlight falls across her in a sharp diagonal, with deep soft shadows. Shot from slightly above, natural skin texture with freckles, light film grain, faded warm colors. A small white serif caption along the bottom edge reads: "Sunday, somewhere around four."

Quick answer

What running H3 Image takes

The H3 Image requirements in eight lines, from MiniMax's model card and Hugging Face, checked on November 11, 2026. We have not run any of these files on our own hardware.

Official model
68.65 GB in BF16
33.1 billion parameters, in MiniMaxAI/MiniMax-H3-Image.
Text encoder
66.7 GB, a separate download
From the text_encoder folder of MiniMaxAI/MiniMax-H3.
Minimum graphics card
Not listed by MiniMax
For vLLM-Omni it names the NVIDIA H200, RTX PRO 6000 Blackwell and AMD Instinct MI350X.
Smallest community file
6.9 GB (GGUF Q2_K)
A community quantization of a pruned model; the text encoder and VAE come on top.
Smallest complete community set
About 21 GB
Comfy-Org's ComfyUI repack: the W4A8 model (12.5 GB), its smallest text encoder (7.3 GB) and the VAE (1.2 GB).
Tools
ComfyUI, Diffusers, SGLang, vLLM-Omni
Settings
6 steps, guidance 1, no negative prompt
License
MiniMax-H3-Image Community License
Does not apply in the US, the EU, the UK or South Korea.

MiniMax H3 Image model card on Hugging Face

H3 Image download: the three parts you need

The official H3 Image download is the image model itself: a 33.1-billion-parameter diffusion transformer in BF16, 68.65 GB in total, in MiniMax's official H3 Image Hugging Face repo, MiniMaxAI/MiniMax-H3-Image. It does not include the text encoder, the part that reads your prompt. H3 Image uses MiniMax H3's Qwen3-VL-based encoder, a separate 66.7 GB download from the text_encoder folder of the MiniMaxAI/MiniMax-H3 repo, so the full-precision setup is about 135 GB.

The third part is the VAE, which turns the model's compressed output into the final picture; H3 Image's VAE downsamples 16 times, and where a community bundle includes it, it is about 1.2 GB. MiniMax also publishes an optional prompt expander, MiniMax-H3-Image-PE (17.9 GB in NVFP4), which rewrites short prompts into the detailed descriptions the model prefers; skip it if you write detailed prompts yourself.

As of November 11, 2026

Every H3 Image Hugging Face version and its size

The official release and the community versions on Hugging Face as of November 11, 2026, with sizes from Hugging Face's file listings. Community versions are uploads by independent people and teams, not MiniMax, and we have not tested their quality. Pruned means parts of the model were removed to make it smaller.

MiniMaxAI/MiniMax-H3-ImageOfficial, by MiniMaxThe H3 Image model in BF16, 33.1B parameters68.65 GBThe 66.7 GB text encoder from MiniMaxAI/MiniMax-H3
MiniMax-H3-Image-PEOfficial prompt expander, optionalRewrites short prompts into detailed ones, in NVFP417.9 GBThe official model; it only rewrites prompts
Comfy-Org/MiniMax-H3-ImageComfyUI repackThe ComfyUI repack: pruned model files in BF16, INT8 or W4A8, three text encoder choices and the VAEModel 40.2 GB (BF16) · 21.0 GB (INT8) · 12.5 GB (W4A8); text encoder 27.1, 15.7 or 7.3 GB; VAE 1.2 GB. The smallest set is about 21 GBNothing else: pick one model file, one text encoder and the VAE from the same repo
Abiray/MiniMax-H3-Image-GGUFCommunity GGUFA pruned model in Q2_K, Q3_K_M, Q4_K_M and Q5_K_M6.9 GB (Q2_K) · 8.9 GB (Q3_K_M) · 11.6 GB (Q4_K_M) · 14.1 GB (Q5_K_M)A text encoder and the VAE: the files are the diffusion model only
unsloth/MiniMax-H3-Image-FP8Community FP8 and INT8The model in FP8 or INT820.3 GB eachA text encoder and the VAE: the repo holds the diffusion model only
Woovee/MiniMax-H3-Image-LowVRAMCommunity low-VRAM bundlePruned W4A8 model, NVFP4 text encoder and VAE together29.4 GB in total (12.5 + 15.7 + 1.2 GB)Nothing listed: all three parts are in the bundle
cguldogan/MiniMax-H3-Image-MLX-4bit and -8bitCommunity MLX for Apple Silicon4-bit and 8-bit versions for MLX31.4 GB in total for 4-bit · 53.3 GB in total for 8-bitAn Apple Silicon Mac with MLX
VisualInference/MiniMax-H3-Image-MLX-bf16Community MLX, full precisionA bf16 version for MLX94.2 GB in totalAn Apple Silicon Mac with MLX
Other quantizationsSD-Next SDNQ uint4, LSDJesus fp8-nvfp4, nativ-community, WaveCut OrbitQuantMore quantization formatsVaries by repoVaries by repo
akhaliq/minimax-h3-image-pdmd-2stepCommunity 2-step adapterA 2-step version, shipped as a LoRA, full BF16 weights and a ComfyUI fileAbout 110 GB for the whole repo; you download only the files you useFor the LoRA: a copy of the model to load it onto

H3 Image VRAM: what the sizes do and don't tell you

Download size is not the H3 Image VRAM you need. Besides the model, the text encoder and the VAE also need memory while a picture is made, and H3 Image renders every picture at 2048 × 2048 or larger before resizing it, so asking for a smaller picture does not shrink that render. Some tools can move parts of the model to system RAM when graphics memory runs out, which lets a big model run on a smaller card but makes each picture slower.

MiniMax publishes no minimum graphics card, and the hardware it names is data-center and workstation class. For an H3 Image low-VRAM setup, look at the community versions: an H3 Image GGUF starts at 6.9 GB for the model alone, an H3 Image FP8 file is 20.3 GB, the smallest full set in Comfy-Org's ComfyUI repack comes to about 21 GB, and Woovee's bundle puts all three parts in 29.4 GB. Those are download sizes, not a promise that a card of that size will work, and results may differ from the official model in ways we have not measured.

Pruna's P-H3-Image, a separate model fine-tuned from MiniMax H3, publishes a home-card figure: its 4-bit ComfyUI files run in about 21 GB of VRAM. If your machine is far from these numbers, the generator on this site runs the official H3 Image in your browser.

Tools

Which tools run H3 Image

MiniMax documents four ways to run the official weights, including splitting one job across several GPUs; the community has added Mac versions.

  • H3 Image ComfyUI templates

    ComfyUI has official MiniMax H3 Image templates for text to image and for edits, with a step-by-step tutorial on docs.comfy.org, and Comfy-Org hosts a ComfyUI repack of the model on Hugging Face.

  • Diffusers in Python

    In a Python script, load the official repo MiniMaxAI/MiniMax-H3-Image with Diffusers' DiffusionPipeline. Diffusers also runs on Apple's mps backend.

  • SGLang server

    Serve the model behind your own API with sglang serve --model-path MiniMaxAI/MiniMax-H3-Image, for apps and batch jobs.

  • vLLM-Omni server

    Start it with vllm serve MiniMaxAI/MiniMax-H3-Image --omni. MiniMax's named hardware (NVIDIA H200, RTX PRO 6000 Blackwell, AMD Instinct MI350X) is for this setup.

  • MLX on a Mac

    Community MLX versions run H3 Image on Apple Silicon; the Mac section below lists them.

Setup

How to set up H3 Image locally

The same four steps apply whichever tool you use to run H3 Image locally.

  1. 1

    Check your memory first

    Compare your graphics memory, and system RAM if your tool can offload, with the table above; the model, text encoder and VAE all need room.

  2. 2

    Download the model and the text encoder

    Get the H3 Image weights from MiniMaxAI/MiniMax-H3-Image or a community version, and the text encoder from the text_encoder folder of MiniMaxAI/MiniMax-H3 unless your version bundles one.

  3. 3

    Load it in your tool

    Open ComfyUI's H3 Image text-to-image or edit template, or point Diffusers, SGLang or vLLM-Omni at the repo.

  4. 4

    Keep 6 steps and guidance 1

    Leave the negative prompt empty and write a detailed English prompt; when a picture misses, change the prompt, not the sampler.

Settings

H3 Image settings to keep in any tool

MiniMax distilled H3 Image to a fixed recipe. These values come from its model card and ComfyUI's tutorial.

Sampling steps
6
MiniMax's servers reject other step and guidance values.
Guidance
1
Negative prompt
None
Say what you want in the picture instead of listing what to avoid.
Prompt
Detailed English, up to 4,096 tokens
MiniMax's GitHub repo includes a prompt-writing guide, skills/h3-image-prompt-writing.
Output size
About 16 megapixels at every aspect ratio
The model works from about 4 to 20 megapixels; by default it renders at least 2048 × 2048, then resizes.
Reference images
Up to 16 for edits
The model card's own code names them in the prompt as <Picture 1>, <Picture 2> and so on.

ComfyUI's MiniMax H3 Image tutorial

H3 Image Mac options: MLX builds and Diffusers

For an H3 Image Mac setup, look at the community H3 Image MLX builds for Apple Silicon: cguldogan's 4-bit and 8-bit uploads and VisualInference's bf16 version. The 4-bit set is 31.4 GB in total, the smallest complete Mac version we found on November 11, 2026. Whether one fits depends on your Mac's memory, and we have not tested any of them. Diffusers can also run the official pipeline on Apple's mps backend, at full size plus the 66.7 GB text encoder. These community conversions are not MiniMax releases, and if your Mac has less memory than the download, the browser generator here runs the official model.

cguldogan's MLX 4-bit build on Hugging Face

H3 Image license: what to check before you build on it

The H3 Image license for running the weights yourself is the MiniMax-H3-Image Community License, dated November 10, 2026. It is royalty-free, but only inside what it calls the Applicable Territory: it excludes the European Union, the United Kingdom, South Korea and the United States, and the exclusion covers using the outputs there as well as running the weights.

Two cases need MiniMax's written permission first: a business with annualized revenue above 1 million US dollars, and anyone offering the model to third parties as a hosted service. You request it through platform.minimax.io/license. A commercial product built on the weights must display the name "MiniMax-H3-Image", and neither the model nor its outputs may be used to improve another AI model, for example as training data. MiniMax claims no rights over the pictures you make, and the license is governed by Hong Kong law.

MiniMax-H3-Image license on GitHub

Where to try H3 Image without a GPU

MiniMax runs an official Hugging Face Space, MiniMaxAI/minimax-h3-image, and two community Spaces offer variants: akhaliq's 2-step version and multimodalart's MiniMax-H3-images. H3 Image is also in MiniMax's Hailuo app, and developers can call it through the MiniMax API.

On this site, the generator runs the official MiniMax H3 Image for text to image, and for edits with up to 16 reference photos. In our runs on November 11, 2026, a 2K picture took 15 to 19 seconds, 4K about 22 seconds and an edit 20 to 24 seconds. A new account can make five 1K pictures free, so you can see real results before you commit to a 135 GB download.

MiniMax's official H3 Image Space on Hugging Face

Run H3 Image Locally examples

  • H3 Image

    A bright phone snapshot of a smiling woman in her mid-twenties leaning on the tiled edge of a rooftop swimming pool, arms folded on the edge, one hand making a peace sign near her face. Wet dark hair in a loose braid over one shoulder, a pale yellow one-piece swimsuit, water droplets on her skin and a light sunburn blush on her cheeks. Behind her: blue pool water, white sun loungers, two palm trees and a hazy city skyline under a cloudy summer sky. Midday sun, slightly overexposed highlights, eye-level framing a little off center, natural skin texture, casual and happy.

  • H3 Image

    A mirror selfie in a minimal apartment hallway: a young woman in her twenties holds a pale blue phone in front of her face so her face is hidden, long straight black hair falling past her shoulders. She wears a fitted white long-sleeve cropped top, light blue striped pajama pants and a small cream teddy-bear bag on a thin gold chain over her shoulder. Soft daylight from a doorway on the left, a white wall and a dark wooden door frame behind her. Slightly tilted vertical framing, phone camera look, true-to-life colors.

  • H3 Image

    An ordinary phone photo taken by a friend across the table in a busy cafe: a woman in her late twenties with messy dark hair and a gray knit sweater, mid-laugh, holding a paper coffee cup. Eye level, slightly tilted framing, her face a little off center. Mixed light from the window and warm ceiling bulbs, a bit of motion blur on her hand, other customers soft in the background. No studio lighting, no retouching, visible skin texture.

  • H3 Image

    Character reference sheet of one young woman with short silver hair, a green bomber jacket, black cargo pants and white sneakers. Top row: full-body front view, side view and back view in the same neutral pose, evenly spaced on a plain light gray background. Bottom row: six head-and-shoulder close-ups with different expressions, labeled Happy, Angry, Surprised, Sad, Thinking and Laughing. Clean digital illustration, flat even lighting, the same proportions and colors in every view.

  • H3 Image

    Knolling photo shot from directly above: a vintage film camera taken apart into its body, lens, lens cap, strap, two film rolls, a cleaning brush and a small screwdriver, all laid out at right angles with even gaps on a pale oak table. Soft daylight from a window on the left, gentle shadows, every object fully visible and none overlapping. Sharp focus from edge to edge, natural colors.

FAQ

Running H3 Image locally: questions