MiniMax H3-Image: 4K Images in 10 Seconds, With One Big Catch

MiniMax launched H3-Image on November 10, 2026 — an image model built on its H3 video system that generates a 4K image, 4096×4096 pixels, in about ten seconds, with open weights. Sixteen reference images in a single edit. A 50% discount for launch week. On paper, that's the most interesting open image model of the year.

Then I read the license. That's where this story actually lives.

I haven't run H3-Image myself. Nobody outside MiniMax has had the weights for more than a day. This article is built from MiniMax's announcement, the model card, and independent launch coverage, checked 11 November 2026. Where a number is MiniMax's, I'll say so — because a launch-day claim is a vendor claim, not an independent result.

What H3-Image actually does

The headline spec is resolution without an upscaler in the middle. H3-Image generates natively at around 16 megapixels at every aspect ratio — 4096×4096 at 1:1, 5472×3072 at 16:9, 6272×2688 at 21:9 — with the API capping output at 20 megapixels. A 4096×4096 image holds about 16.8 million pixels, more than double a 4K television screen. That's print-plate territory, or a background you can cut three placements out of without it falling apart.

The second spec is the one I'd actually build on: up to 16 reference images in one request, addressed positionally as through in upload order. Text-to-image and editing run through one endpoint pair, so the same call that generates a scene can also edit it using your references — a character from one photo, a style from another, a layout from a third. Character consistency across a series is the thing image models usually fumble. Sixteen reference slots is a serious attempt at fixing it.

It's available three ways: the MiniMax API under the model id MiniMax-H3-Image, MCode, and Hailuo AI. The weights are on Hugging Face at MiniMaxAI/MiniMax-H3-Image — though at announcement time the link was unreachable for several users, so if you hit a dead page yesterday, that was everyone, not just you.

The speed claim, honestly

MiniMax's blog says a full 4K image lands in under 10 seconds, "at up to 5× the speed of mainstream alternatives" — Standard mode 1.5–2× as fast, Highspeed mode 3–4×. Launch coverage of the hosted service puts it more conservatively at 10–20 seconds. Either way, that's fast for 16 megapixels. But read it as the hosted service's time, not the time on your own GPU. Nobody has published an independent benchmark yet: as of November 11, H3-Image has no entry on either of Artificial Analysis's image boards, and the speed and quality figures are all MiniMax's own.

The only independent test I've seen so far is Pixel Dojo running the same eight prompts on H3-Image and GPT Image 2.6, one try each. H3-Image made the more natural-looking photo; GPT Image 2.6 came out more stylised and liked adding things of its own. Eight prompts, one try each, is a sniff test, not a verdict — but it's the only non-vendor data point that exists right now, and it points in an encouraging direction.

Using it via the API

The practical detail that matters if you're wiring this in today: the served model is distilled to six sampling steps, and those two parameters are fixed. Pass anything else and the API returns an error:

{
  "model": "MiniMax-H3-Image",
  "num_inference_steps": 6,
  "guidance_scale": 1
}

That fixed sampler is actually good news for reproducibility — the same prompt and references should give you the same result every time, which is not something you can say about most image APIs. Check MiniMax's API reference for the exact endpoint shape before you build; the model id and the fixed parameters are the stable parts.

On price: the announcement itself gives no per-image base price. What MiniMax does say is that pricing runs at 50–70% of comparable mainstream offerings, and that launch week brings 50% off the standard price for generations at 4K and above. Per MiniMax's rate card, that discount window runs to 17 November, 16:00 UTC, and only applies to images of 16.6 megapixels and above. A discount on an unknown base rate is a discount you can't do math on yet — check the rate card before you promise a client a number.

Self-hosting: bring a serious GPU

Open weights does not mean lightweight. H3-Image is a 33.1-billion-parameter dense diffusion transformer, with about 68.65 GB of weights plus a 66.7 GB text encoder borrowed from the H3 video model. It deploys through SGLang, vLLM-Omni, diffusers and ComfyUI, but this is datacentre-or-a-very-serious-workstation territory, not a laptop job. If your plan was "download it tonight and run it locally," budget for the hardware first.

The catch, stated plainly

• The license may ban half the audience. MiniMax's announcement doesn't name H3-Image's own license. The base H3 model ships under the MiniMax H3 Community License, which permits use only outside the European Union, the United Kingdom, South Korea and the United States. If that license carries over to H3-Image — and nobody has confirmed otherwise yet — developers in the US and EU legally can't run the weights. The hosted API may sit under different terms, but check before you build anything on it. This is the single most important fact in the launch, and it's the one the announcement doesn't spell out.
• No base price in the announcement. A 50% discount sounds great until you realise the starting rate isn't in the launch post. MiniMax's "50–70% of mainstream offerings" is a claim, not a price list.
• Every speed and quality number is MiniMax's. No independent benchmark exists yet, no Artificial Analysis ranking, and the one third-party test so far is eight prompts. The 10-second claim is plausible; it's also unverified.
• No transparent backgrounds. Per MiniMax's own API reference, transparent backgrounds aren't supported — so if your workflow needs real alpha channels, this isn't your model. ByteDance's Seedream 5.5, announced a day earlier on November 9, does offer them, at 2K.
• The Hugging Face link was down at launch. Several users couldn't reach it when the announcement went out. It works now, but a launch where the weights link doesn't open on day one is worth noting.

The bottom line

H3-Image is a genuine step: 16-megapixel images in about ten seconds, sixteen reference images in one edit, open weights, and a price that undercuts the mainstream — if the discount and the rate card hold up. For print-ready assets, multi-reference character work, and anyone who wants a reproducible six-step sampler they can run themselves, it's the most practical open image launch of the year so far.

But the license question comes first, not last. Before you download 68 GB of weights or build a feature on the API, find out whether you're even allowed to use them where you live. A model this good with a license this narrow is a story about access, not just resolution.

If you're in the US or the EU, does the license kill H3-Image for you before the specs even matter — or is the hosted API enough?

Sources: MiniMax, "MiniMax H3-Image: Bringing 4K Image Generation into the 10-Second Era" (https://www.minimax.io/blog/minimax-h3-image), 10 November 2026; Hugging Face model page (https://huggingface.co/MiniMaxAI/MiniMax-H3-Image); PANews, "MiniMax Open-Sources MiniMax-H3-Image Model" (https://www.panews.io/articles/01a1c625-9021-743f-960d-733355662823), 10 November 2026; Pixel Dojo, "MiniMax H3 Image vs GPT Image 2.6" (https://pixeldojo.ai/guides/minimax-h3-image-vs-gpt-image-2-6); independent launch coverage (https://pasqualepillitteri.it/en/news/31797/minimax-h3-image-4k-open-weights-launch). All capability, speed, pricing and benchmark claims are MiniMax's unless stated otherwise. No hands-on testing was performed for this article. Nothing here is sponsored — no affiliate links on this page. This breaking article is not the daily buyer-intent post. Plain link for readers: minimax.io.