Skip to content

AI Tools

What Is MiniMax H3-Image? Features, Pricing & How to Use It

Quick Answer

MiniMax H3-Image is MiniMax’s image generation and editing model, released November 10, 2026. It generates and edits images at about 16 megapixels natively (4096×4096 at square format, up to 6272×2688 at 21:9), takes up to 16 reference images, and is built on a 33.1-billion-parameter diffusion transformer. You can use it in the Hailuo AI app, through the MiniMax API, in MiniMax Code, or by running the open weights yourself — if you are outside the four excluded regions. On the official API a 4K image costs about $0.069, or about $0.034 during the launch offer until November 17, 2026. It cannot make transparent backgrounds, and it is weak at text with diacritics (accented letters).

What Is MiniMax H3-Image?

MiniMax H3-Image is the image model of MiniMax’s H3 family. MiniMax H3 is the company’s video model, which makes clips with native sound; H3-Image uses the same architecture with the audio branch removed, so it only makes images. One set of weights handles both jobs: text-to-image generation and editing from reference images.

It is built on a 33.1-billion-parameter diffusion transformer (DiT), with a text encoder taken from MiniMax H3 (based on Qwen3-VL). The model is distilled to run in just six sampling steps, with no classifier-free guidance and no negative prompt — the API rejects any other step count or guidance value. There is also an optional prompt expander, MiniMax-H3-Image-PE, which turns a short prompt in any language into a long English description.

MiniMax is the company behind Hailuo AI, its consumer app, and the H3 video model. H3-Image is available through the MiniMax API (platform.minimax.io), the Hailuo AI app, MiniMax Code, and as open weights on Hugging Face. Third-party hosted options such as fal.ai and Venice also carry it.

MiniMax H3-Image Release Date and Availability

MiniMax announced H3-Image on November 10, 2026, in a post on X: “16 million pixel. 10-20 seconds. Open weights.” The weights went public on Hugging Face the same day, in the MiniMaxAI/MiniMax-H3-Image repository — the repository had been created on November 3 and sat empty until the weights landed on November 10.

It is available in:

  • Hailuo AI app — the easiest way to try it, with free starter credits for new users.
  • MiniMax API — platform.minimax.io, pay-as-you-go by token, for developers.
  • MiniMax Code — MiniMax’s coding agent environment.
  • Hugging Face — open weights for self-hosting, under the MiniMax-H3-Image Community License.
  • Third-party hosts — fal.ai and Venice, among others.

MiniMax H3-Image Features

1. Native 16-megapixel output, no upscaler

Most image models generate a smaller picture and enlarge it afterwards. H3-Image draws about 16 megapixels directly, at any aspect ratio: 4096×4096 at 1:1, 4736×3552 at 4:3, 5024×3360 at 3:2, 5472×3072 at 16:9, and 6272×2688 at 21:9. The working range is roughly 4 to 20 megapixels, and the model card says results are best at around 16 megapixels. That makes it useful for posters and large-format print without a second upscaling step.

2. Up to 16 reference images

Editing and composition accept up to 16 reference images in a single request — far more than most rivals. You refer to them positionally in the prompt, as <Picture 1> through <Picture 16>, in upload order. In one independent test, a living room composed from 16 references placed 15 of the items in the scene. This is the model’s standout feature for product shots, interiors, and scenes built from many objects.

3. Generation and editing in one model

The same checkpoint does text-to-image generation and image editing. Each API request can generate 1 to 10 images, and outputs can be JPEG, PNG or WebP, up to 20 megapixels. Editing works by describing the change and the references; there is no mask-based editing, no transparent-background option, and no image preview in the API.

4. Fast by design — with a caveat

MiniMax says a 4096×4096 image takes “under 10 seconds”. That is a vendor-reported figure. MiniMax’s own comparison chart shows a median of 16.0 seconds in Standard mode and 10.3 seconds in Highspeed mode for 4K text-to-image, and one independent tester measured 13–17 seconds at 2K and 22–24 seconds at 4K through the API. Treat the 10-second figure as a best case on MiniMax’s own infrastructure, not a typical API time.

5. A Highspeed variant

There is a faster model, MiniMax-H3-Image-Highspeed, with the same parameters and limits. It costs twice as much per image token as the standard model (text input tokens cost the same). In one independent test it was not actually faster at 2K, so it is worth testing before paying double.

6. Open weights — with limits

The weights are public on Hugging Face: 33.1 billion parameters, about 68.65 GB for the transformer and VAE. You also need a separate 66.7 GB text encoder from MiniMax-H3, so a first local run means downloading about 135 GB. The community has already published smaller ComfyUI versions (int8 at 21 GB, w4a8 at 12.6 GB) and GGUF files (from 6.9 GB). Recommended runners include SGLang, vLLM-Omni, diffusers and ComfyUI. The licence, however, is the part to read before downloading — see Limitations below.

MiniMax H3-Image Pricing Explained

The official MiniMax API bills by token, not per image. The rates are:

Billing item Price per 1M tokens
Text input $2.30
Image input $1.50
Image output (Standard) $3.80
Image output (Highspeed) $7.60

MiniMax’s own worked examples put a standard text-to-image picture at about $0.022 for 1K, $0.023 for 2K, and $0.069 for 4K. A 4K edit with one reference image costs about $0.037 at the launch rate. Your bill depends on the number of reference images, the dimensions, and the text input.

Launch offer: until November 17, 2026, 16:00 UTC, images of at least 16.6 million pixels cost half price on the API — a 4K image is about $0.034 instead of $0.069. The offer halves the cost of the whole request, for both Standard and Highspeed. Watch the threshold: a 3648×4544 portrait (16.58 megapixels) falls just under it and was charged the full price in one test, while 4096×4096 and 5472×3072 qualified.

Hailuo AI app: the app sells monthly credit plans — reported in November 2026 from about $10.50 a month for 1,000 credits, up to $216 a month for 27,000 credits — and new users get free starter credits. One H3-Image picture costs a few credits depending on resolution; check the current rate inside the app before buying credits, as reported figures differ. A 4K image works out to roughly $0.05 on the higher plans, against $0.069 on MiniMax’s own API. The same Hailuo credits also run other models, including Midjourney V7, Nano Banana Pro, GPT Image 2 and Seedream.

MiniMax H3-Image vs Seedream 5.5

The closest launch to compare with is ByteDance’s Seedream 5.5, released one day earlier, on November 9, 2026. Both models pitch themselves on editing. The comparison is thin because ByteDance has not disclosed Seedream 5.5’s key specs.

Feature MiniMax H3-Image Seedream 5.5
Release date November 10, 2026 November 9, 2026
Native output resolution About 16 MP, no upscaler Not officially disclosed
Reference images Up to 16 Not officially disclosed
Output area limit 20 megapixels Not officially disclosed
Open weights Yes (community licence) No
Transparent backgrounds No Yes (RGBA, in one independent test)

Against the other big image models, the picture from one independent tester’s launch-day comparison (same prompts, November 9–10, 2026) is: H3-Image is the cheap 4K option and the best at composing many references, while Nano Banana 2.1 and GPT-Image-2.6 handle text with diacritics far better, and GPT-Image-2.6 held a person’s likeness better in a face test. For a broader look at the options, see our 10 Best AI Image Generators in 2026 guide.

How to Use MiniMax H3-Image

In the Hailuo AI app (easiest)

  1. Open the Hailuo AI app (hailuoai.video) and sign in. New users get free starter credits.
  2. Open the image creation panel and pick H3-Image (or H3-Image Highspeed).
  3. Write your prompt — prompts can be up to 7,500 characters, and you can write in any language; the prompt expander rewrites it into English.
  4. To edit or compose, upload up to 16 reference images and name them in the prompt as <Picture 1>, <Picture 2>, and so on.
  5. Choose the size (1K, 2K or 4K) and generate. Download the result as JPEG, PNG or WebP.

Through the MiniMax API (developers)

The API has two endpoints: POST /v1/images/generations for text-to-image and POST /v1/images/edits for editing with references. The model names are MiniMax-H3-Image and MiniMax-H3-Image-Highspeed. A request can include 1–16 reference images and return 1–10 images. Two practical notes from the API guide: the server keeps generating after a client timeout or disconnect and still bills for successful images, so disable automatic SDK retries (max_retries=0); and the API accepts only 6 inference steps with a guidance scale of 1 — anything else returns an error, and negative prompts are refused.

Running the weights yourself

Download the weights from the MiniMaxAI/MiniMax-H3-Image repository on Hugging Face, plus the text encoder from MiniMax-H3 (about 135 GB together). Run them with SGLang, vLLM-Omni, diffusers or ComfyUI. Check the licence first: if you are in the EU, UK, US or South Korea, the community licence does not cover you.

Best Uses for MiniMax H3-Image

  • Product and e-commerce photos: place a product in a new scene while keeping its label and shape, using the original photo as a reference.
  • Covers and posters with English text: book covers, thumbnails and posters came out clean in independent tests, including titles and author names.
  • Storyboards: a six-panel storyboard sheet with numbered panels and camera arrows in one image.
  • Illustrations in many styles: pencil, flat illustration, macro photography, wildlife — the model handles varied styles in one checkpoint.
  • Interiors from many references: compose a room from up to 16 items at once.
  • Large-format print: 16 megapixels straight from the model, with no upscaling step.

For free alternatives, see our Best Free AI Image Generators in 2026 guide.

Limitations and Things to Know

  • The licence excludes four regions. The MiniMax-H3-Image Community License, dated November 10, 2026, covers the whole world except the European Union, the United Kingdom, South Korea and the United States. Use, reproduction, modification, distribution or display of the weights or their outputs is not authorized there — people in those places must contact MiniMax for a separate licence. Running the model in a permitted country does not grant the right to display its output elsewhere.
  • Revenue and hosting gates. Companies with annualized revenue above $1 million (including affiliates) need MiniMax’s written authorization, and so does anyone offering the model as a hosted service, with no revenue threshold. Commercial products using the model must display the name “MiniMax-H3-Image” in their interface. Outputs may not be used to train or improve other AI models.
  • It is “open weights”, not open source. The files are public, but the licence is MiniMax’s own — not MIT or Apache 2.0. Hugging Face lists it simply as “license: other”. MiniMax claims no rights to generated outputs.
  • No transparent backgrounds. The API does not support them, and in one independent test 0 of 5 images came back with a transparent background.
  • Text with diacritics is weak. English text inside images was clean in independent tests, but Polish text with diacritics averaged 26 errors per 550 characters. Nano Banana 2.1 and GPT-Image-2.6 did far better on the same test.
  • Chained edits build up texture. Editing the same image round after round works, but visible texture builds up from round 2–3. Plan edits in as few rounds as possible.
  • No masks or previews in the API. Masked edits, transparent backgrounds and image previews are all unsupported on the hosted API.
  • No benchmarks published. MiniMax has published no benchmark scores for H3-Image, and no independent benchmark exists yet. Quality claims rest on MiniMax’s own demos and early hands-on tests.
  • The 10-second claim needs context. MiniMax’s own chart shows 16.0 seconds (Standard) and 10.3 seconds (Highspeed) as medians for 4K text-to-image, and an independent API test measured 22–24 seconds at 4K. Real times depend on the mode, the queue and the infrastructure.
  • Prices and limits can change. The launch offer ends November 17, 2026, and Hailuo credit rates differ by plan and resolution — check the live price before planning a batch.

Frequently Asked Questions

What is MiniMax H3-Image?

An image generation and editing model from MiniMax, released November 10, 2026. It makes native 16-megapixel images and edits from up to 16 reference images, built on a 33.1-billion-parameter diffusion transformer.

When was MiniMax H3-Image released?

November 10, 2026, announced on X and published the same day on Hugging Face, the MiniMax API, Hailuo AI and MiniMax Code.

How much does MiniMax H3-Image cost?

On the official API, about $0.023 for a 2K image and $0.069 for a 4K image, or about $0.034 for 4K during the launch offer until November 17, 2026. In the Hailuo app it costs a few credits per image on a monthly plan.

Is MiniMax H3-Image free?

Not entirely. Hailuo AI gives new users free starter credits, so you can try it without paying. The MiniMax API charges from the first image, and self-hosting the weights costs nothing in licence fees — if the licence covers you.

Where can I use MiniMax H3-Image?

In the Hailuo AI app, through the MiniMax API, in MiniMax Code, by running the open weights from Hugging Face, or through third-party hosts such as fal.ai and Venice.

Can I run the weights myself?

Yes, if you are outside the EU, UK, US and South Korea. You need about 135 GB of downloads (68.65 GB of model weights plus a 66.7 GB text encoder) and a runner such as SGLang, vLLM-Omni, diffusers or ComfyUI. Companies above $1 million in annual revenue need MiniMax’s written authorization first.

Can MiniMax H3-Image make transparent backgrounds?

No. Transparent backgrounds are not supported by the API, and independent tests returned full backgrounds every time.

How is MiniMax H3-Image different from MiniMax H3?

MiniMax H3 is a video model with native sound. H3-Image uses the same architecture without the audio branch and only makes images. For video options, see our 10 Best AI Video Generators in 2026 guide.

Final Verdict

MiniMax H3-Image is the best-value 4K image model of its launch week: native 16-megapixel output, up to 16 references in one picture, and about $0.034 a 4K image until November 17. For product photos, posters, storyboards and scenes built from many references, it is a strong pick — especially if you want to self-host. The honest catch is the licence: the open weights are closed to the US, EU, UK and South Korea, and the model cannot do transparent backgrounds or accented text well. If you are in a covered region and need cheap 4K, use the API or Hailuo instead of the weights; if you need cut-outs or diacritics, pick a different model.

Editorial update: November 12, 2026. This article is based on MiniMax’s launch announcement, the Hugging Face model card, MiniMax’s API pricing page, and independent launch-day tests, checked November 12, 2026.