Last updated: November 15, 2026
MiniMax H3-Image is an open-weight image model that generates and edits pictures at about 16 megapixels in six sampling steps. MiniMax released it on November 10, 2026. You can download the weights, run it in ComfyUI, SGLang or vLLM-Omni, or call it through the MiniMax API.
This guide draws on MiniMax’s launch post, the official GitHub repository, the community license, the MiniMax pricing page and the text-to-image API reference. For MiniMax’s latest language model, see our MiniMax M3.1 Flash guide.
Contents
- What is new
- Specifications at a glance
- API pricing
- How to run it locally
- Prompting tips
- License limits
- How it compares
- Frequently asked questions
What is new
The license gives November 10, 2026 as the release date, and MiniMax’s blog post is dated November 11. H3-Image is built on MiniMax H3, the company’s open-weight video model, and follows the H3-Base design with the audio branch removed.
MiniMax makes three main claims. It says H3-Image can produce a 4096 × 4096 image in under 10 seconds, at up to five times the speed of mainstream alternatives. It says the images avoid the overly smooth look common in AI output. It also says the model has the strongest prompt adherence among open-weight models. These are vendor claims and have not been independently verified.
The model is also available in the MiniMax API Platform, MiniMax Code and Hailuo AI. According to the blog, the Standard mode is 1.5 to 2 times as fast as mainstream alternatives and the Highspeed mode is 3 to 4 times as fast.
Specifications at a glance
The repository lists these specifications for the API and the model:
| Item | MiniMax H3-Image |
|---|---|
| API model IDs | MiniMax-H3-Image, MiniMax-H3-Image-Highspeed |
| Tasks | Text-to-image and image editing in one checkpoint |
| Architecture | 33.1B-parameter dense single-stream diffusion transformer, 50 layers; Qwen3-VL text encoder from MiniMax-H3 |
| Output size | 1K, 2K or 4K area; custom sizes from 655,360 to 20,000,000 pixels |
| Aspect ratio | 1:3 to 3:1 |
| Prompt | Up to 32,000 characters; content beyond 4,096 tokens is truncated |
| Reference images | Up to 16, each up to 64 MiB, JPG, PNG, WebP, HEIC or HEIF |
| Images per request | 1 to 10 |
| Not yet supported | Mask-based inpainting, transparent backgrounds, user-set seeds (API), preview images |
Prompt expansion is on by default for text-to-image through the API and can be turned off with disable_prompt_enhanced. It is always on for image editing.
API pricing
MiniMax bills H3-Image per million tokens. The pricing page lists:
| Model | Input text | Input image | Output image | Example: 1K text-to-image | Example: 4K text-to-image |
|---|---|---|---|---|---|
| MiniMax-H3-Image (Standard) | USD 2.30/M | USD 1.50/M | USD 3.80/M | USD 0.02179 | USD 0.06879 |
| MiniMax-H3-Image-Highspeed | USD 2.30/M | USD 3.00/M | USD 7.60/M | USD 0.03529 | USD 0.13001 |
Until November 17, 2026 at 16:00 UTC, MiniMax takes 50% off the total request cost when the delivered image has at least 16,600,000 pixels. That brings the 4K Standard example to about USD 0.034. Text input tokens include tokens from prompt enhancement, so actual costs vary.
Two API details matter for budgeting. The API reference says the endpoint is not idempotent: if your client times out, the server still finishes and charges for successful images, and a retry charges again. MiniMax recommends a client timeout of at least 600 seconds and, with the OpenAI SDK, max_retries=0. Only successful images are charged, so check data[].error for images withheld by the content safety check.
How to run it locally
The repository recommends SGLang, vLLM-Omni, diffusers or ComfyUI. A typical SGLang setup:
- Check the license first. Confirm you are outside the excluded territories described below.
- Download the weights. Run
hf download MiniMaxAI/MiniMax-H3-Imagewith the original checkpoint folders, thenhf download MiniMaxAI/MiniMax-H3 --include "text_encoder/*". This needs about 135 GB; the full repository with the diffusers format is about 204 GB. - Install SGLang from the main branch with
pip install -e 'python[diffusion]'. - Start the server with
sglang serve --model-path MiniMaxAI/MiniMax-H3-Image --port 30000. Add--num-gpus 4 --ulysses-degree 4to split requests over four GPUs. - Expand your prompt. A local server uses the prompt as given, so expand it first with the optional MiniMax-H3-Image-PE model (17.9 GB) or the API’s Context-IR step.
- Send a request to
/v1/images/generationswith asizesuch as4096x4096. Leavenum_inference_stepsat 6 andguidance_scaleat 1; other values return HTTP 400.
MiniMax says SGLang offloads components when less than 130 GiB of GPU memory is free, and its cookbook has recipes from B300 down to 12 GB cards. ComfyUI supports the model natively and through API-based Partner Nodes. For background on these servers, see our SGLang 0.5.23 guide and vLLM 0.32 guide.
Prompting tips
The repository summarizes MiniMax’s official prompting advice:
- State the task and content first, then layout, then style through visible color, material and light.
- Quote every word that should appear, with its position. Say “no text” if you want none.
- Give exact data for charts, with units and time ranges.
- Give each reference image a role, and say what changes and what stays.
- Describe grids as rows and columns, such as “2 rows and 3 columns”.
MiniMax also publishes an h3-image-prompt-writing skill for agents such as Claude Code, Codex and Cursor.
License limits
This is the most important caveat. The MiniMax-H3-Image Community License grants rights only in an “Applicable Territory” that excludes the European Union, the United Kingdom, South Korea and the United States. Users there cannot rely on the license to run the weights.
Inside the permitted territory, you need separate written authorization from MiniMax if you or your affiliates exceed USD 1 million in annualized revenue or if you offer the model as a hosted service. You may not use the model or its outputs to improve other AI models. MiniMax claims no rights over the outputs you generate.
How it compares
Within MiniMax’s lineup, the pricing page now lists the earlier image-01 model under legacy models, at USD 0.0035 per image. H3-Image costs more per image but targets much higher resolution and accepts up to 16 reference images.
For a closed-model alternative from ByteDance, see our Seedream 5.5 guide. If you only need occasional images in a chat app, our answer on the ChatGPT image generation limit covers consumer options.
Frequently asked questions
What is MiniMax H3-Image?
MiniMax H3-Image is an open-weight image generation and editing model from MiniMax, built on its H3 model family. One checkpoint handles text-to-image, reference-based generation and image editing, with output of about 16 megapixels in six sampling steps.
Is MiniMax H3-Image free to download?
The weights are published on Hugging Face under the MiniMax-H3-Image Community License. The license is royalty-free but excludes the United States, the European Union, the United Kingdom and South Korea, and it requires written authorization above USD 1 million in annualized revenue or for hosted services.
How much does the MiniMax H3-Image API cost?
MiniMax lists USD 3.80 per million output image tokens for MiniMax-H3-Image and USD 7.60 for MiniMax-H3-Image-Highspeed. Its example costs are about USD 0.022 for a 1K text-to-image result and USD 0.069 for 4K, halved for 4K until November 17, 2026.
What hardware do I need to run H3-Image locally?
The full download is about 204 GB, or about 135 GB for SGLang and vLLM-Omni. MiniMax says one B200 or B300 holds the whole model in about 135 GiB, one H200 works with text encoder offload, and the SGLang cookbook covers cards down to 12 GB.
Can I use H3-Image in ComfyUI?
Yes. MiniMax lists ComfyUI as a supported deployment with text-to-image and edit workflow templates. ComfyUI can run the downloaded weights natively or call the MiniMax API through Partner Nodes.
Sources
- MiniMax H3-Image: Bringing 4K Image Generation into the 10-Second Era, MiniMax (November 11, 2026)
- MiniMax-H3-Image repository, GitHub
- MiniMax-H3-Image Community License Agreement, Hugging Face
- Pricing Overview, MiniMax API Platform
- Text-to-Image API reference, MiniMax API Platform