Type a sentence. Count to ten. Now picture a poster-sized image appearing on your screen, sharp enough to print, made entirely by an AI model you can download and run yourself. That is the promise MiniMax made this week with MiniMax H3-Image, which turns words into 16-megapixel pictures in about ten seconds.
The Chinese AI company announced the model on November 10, and the launch landed with a thud in the open-weight AI world. A 4K image in ten seconds is a headline number that grabs attention. But the real story is more complicated than the headline, because the model arrives with a licence that shuts out half the world.
So what exactly did MiniMax release, how fast is it really, and who is actually allowed to use it? Here is the full picture.
What MiniMax H3-Image Actually Does
At its simplest, H3-Image is an image generator and editor in a single model. You type a text prompt and it produces a picture. You can also hand it up to 16 reference images at once and ask it to edit, combine or restyle them.
The headline figure is resolution. H3-Image generates natively at around 16 megapixels at every aspect ratio, with no separate upscaler in the middle. A square image comes out at 4096 by 4096 pixels, a widescreen 16:9 image at 5472 by 3072, and the widest cinematic ratio, 21:9, at 6272 by 2688.
To put that in perspective, a 4096 by 4096 image holds about 16.8 million pixels. That is more than double a 4K television screen, which shows roughly 8.3 million pixels. In practical terms, the model works across a range of about 4 to 20 megapixels, and MiniMax’s API caps total output area at 20 megapixels.
Speed is the second headline claim. According to MiniMax, a full 4K image lands in about ten seconds. Treat that as a vendor claim for now: no independent test has timed the model yet, and MiniMax’s own model card states no generation time at all.
There are three ways to try it. H3-Image is available through MiniMax’s own API, through MiniMax Code, and through Hailuo AI, the company’s consumer app. The weights are also published on Hugging Face, so anyone with enough hardware can download and run the model themselves.
A 33-Billion-Parameter Engine Distilled to Six Steps
Under the hood, H3-Image is a 33.1-billion-parameter dense diffusion transformer. It has 50 layers, a hidden size of 5,376, 56 attention heads, and a feed-forward width of 14,336. Each transformer token covers a 32 by 32 pixel patch of the image.
The speed trick is distillation. MiniMax squeezed the model down to just six sampling steps, with no classifier-free guidance at all. Most image models need dozens of steps to build a picture gradually, so cutting that to six is what makes the ten-second claim plausible.
The served API enforces the shortcut strictly. It rejects anything other than six inference steps and a guidance scale of one, and it refuses negative prompts outright. If you want a different sampling recipe, you will need to run the weights yourself.
Those weights are heavy. The transformer and VAE together total about 68.65 gigabytes, with the transformer alone weighing 66.28 gigabytes in BF16 precision, plus a separate text encoder. This is not a model for a casual laptop. It needs serious GPU memory, or a hosted API, to run at all.
The Licence Catch: Open Weight Does Not Mean Open Everywhere
Here is where the story gets complicated. MiniMax calls H3-Image open-weight, and the weights are genuinely downloadable. But the licence attached to them is not an open-source licence, and it carries conditions that change who can actually use the model.
Under the MiniMax-H3-Image Community License Agreement, use is permitted only outside the European Union, the United Kingdom, South Korea and the United States. That is a remarkable restriction for a model presented as open. It excludes the two largest AI markets in the world, plus much of Europe and East Asia.
The conditions do not stop there. Annual revenue above one million US dollars needs MiniMax’s prior written authorisation. Turning the model into a hosted service needs authorisation too, with no revenue threshold. Commercial interfaces must display the model name, and outputs may not be used to train other models.
The Hugging Face page does not help clarify matters. It reduces the whole licence to a single word: other. Anyone planning to download the 68 gigabytes of weights or build a product on the API should read the licence first, because the headline “open-weight” does not tell the full story.
The contrast with last week’s open-weight releases is stark. Meituan’s LongCat-2.5, a 1.6-trillion-parameter model released on November 9, arrived under the permissive MIT licence.
The MIT licence allows commercial use, modification and redistribution with almost no strings attached. H3-Image’s licence is a different animal entirely.
What It Costs, and the Launch-Week Discount
MiniMax has not published a simple per-image price in its announcement, but its rate card fills in the numbers. Billing is token-based: 2.30 dollars per million input text tokens, 1.50 dollars per million input image tokens, and 3.80 dollars per million output image tokens.
Translated into plain language, a single 4K image costs about 6.9 US cents at list price. During launch week, MiniMax is offering 50 percent off generations of 16 megapixels and above, which brings that down to roughly 3.4 cents per image. The discount window runs until November 17 at 16:00 UTC.
MiniMax says its pricing runs at 50 to 70 percent of comparable mainstream offerings. That is a deliberate pitch to developers who are currently paying closed-model prices for image generation, and it puts H3-Image in the same value conversation as other budget image APIs.
For now, the value question has no independent answer. As of November 11, H3-Image has no entry on either of Artificial Analysis’s image leaderboards. The only independent test seen so far is a small Pixel Dojo comparison of the same eight prompts on H3-Image and GPT Image 2.6.
How H3-Image Stacks Up Against the Competition
The image generation race has split into two camps this year. On one side are the closed giants: Google’s Gemini image models, OpenAI’s GPT Image family, and Adobe’s Firefly, all polished, all expensive, and all locked behind APIs.
On the other side are the open-weight challengers, led by Chinese labs. H3-Image sits in the second camp, but awkwardly.
Its resolution and speed claims are genuinely competitive, and sixteen reference images in one edit beats most rivals. Yet the licence means the model cannot legally be used in the United States or Europe, home to most of the world’s image-generation developers.
That puts it in a strange position next to its Chinese peers. Alibaba’s Qwen4-Max, launched the same week, is preview-only and also closed, while Meituan’s LongCat-2.5 is open under MIT and can be used anywhere. H3-Image is the most restrictive of the three, despite the same open-weight marketing.
The timing is also telling. H3-Image arrived in the same week Google made its strongest model, Gemini 4 Carbon, generally available, and days after Anthropic launched its Claude office suite. A model can dominate the news for a morning and be old news by dinner.
What Happens Next
The next test for H3-Image is not another launch post. It is the independent benchmarks. Once reviewers run the model on standard prompts, the world will learn whether the ten-second 4K images hold up on quality, prompt following and editing accuracy, or whether speed came at a cost.
The licence question will also need an answer. MiniMax has not explained why the United States, the European Union, the United Kingdom and South Korea are excluded, and it is not clear whether the restriction is a legal precaution, a commercial strategy, or something else entirely.
Developers in those regions who want to experiment will have to use the hosted API, if that is even available to them, or wait.
For everyone else, the model is already there to download. The weights landed on Hugging Face on November 10 at 14:50 UTC, after the repository sat empty since November 3. The model card recommends SGLang, vLLM-Omni, Diffusers and ComfyUI for running it, so the tooling for self-hosting is ready from day one.
H3-Image is a genuine step forward in open-weight image generation: 16-megapixel pictures in about ten seconds, sixteen reference images in one edit, and a price that undercuts the mainstream.
But the licence is the story that matters most. Before you download 68 gigabytes of weights or build a feature on the API, find out whether you are even allowed to use them where you live.
