MiniMax has released an image model that makes a 4K picture in about ten seconds — and then, in the fine print, told a large part of the world it cannot use it. H3-Image, announced on 11 November 2026, is a 33.1-billion-parameter diffusion transformer that generates and edits images natively at around 16 megapixels, accepts up to 16 reference images in a single edit, and ships with open weights. It is also the most interesting licensing story in image generation this year, because the license excludes the United States, the European Union, the United Kingdom and South Korea.
That contradiction — open weights, closed map — is the story. The model itself is a genuine step. The license is the part you need to read before you download 68 GB of weights or build a feature on the API.
What H3-Image actually does
H3-Image is built on MiniMax's H3 video system and combines three jobs in one checkpoint: text-to-image generation, reference-based generation, and image editing. The headline spec is resolution without an upscaler in the middle: it 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 total output area at 20 megapixels. MiniMax says a full 4K image lands in under 10 seconds, up to 5× the speed of mainstream alternatives, from a model distilled to six sampling steps.
The reference-image support is the feature creators will actually feel. Up to 16 reference images in a single edit means you can hold a character, a product, a style and a scene steady across generations — the kind of consistency work that usually takes a separate control pipeline. MiniMax also claims state-of-the-art prompt adherence among open-weight models, and says the model avoids the overly smooth, artificial look that marks most AI images.
It is available three ways: the weights on Hugging Face (MiniMaxAI/MiniMax-H3-Image, a 68.65 GB BF16 checkpoint), through MiniMax's own API under the model ID MiniMax-H3-Image, and inside MiniMax Code and Hailuo AI. Pricing, per MiniMax, runs at 50–70% of comparable mainstream offerings, with 50% off standard pricing for 4K-and-above generations during launch week (until 17 November, 16:00 UTC).
The license is the headline
H3-Image does not ship under a standard open-source license. It uses the MiniMax-H3-Image Community License Agreement, and four terms matter more than any benchmark:
- Territory: the license grant excludes the United States, European Union, United Kingdom and South Korea. If you live or operate in one of those places, the weights are downloadable but not licensed to you.
- Revenue: if your annualised revenue exceeds $1 million, you need MiniMax's prior written authorisation.
- Hosting: providing the model as a hosted service to third parties also requires prior written authorisation, with no revenue threshold.
- Outputs: commercial interfaces must prominently display the model name, and outputs may not be used to train or improve other models.
That is why "open weights" and "open source" are not the same sentence. The weights are public, the model card lists Diffusers, SGLang, vLLM-Omni and ComfyUI as deployment options, and one tester reportedly rendered 16-megapixel images on a single 16 GB RTX 5080. But the permission to use what you downloaded stops at a map. For readers in Egypt and most of the world outside the excluded territories, the license is workable; for anyone in the US or EU — the two largest creator markets — it is a wall, and building a product on H3-Image there means asking MiniMax first.
What it cannot do yet — and what nobody has verified
Two honest caveats, both from MiniMax's own materials. The model's GitHub README lists what is missing: no mask-based inpainting, no transparent backgrounds, no user-specified seeds, and no preview images. Those are the tools a working designer reaches for daily, and they are not here yet.
And the speed and quality figures are, for now, entirely MiniMax's. As of 11 November, H3-Image had no entry on either of Artificial Analysis's image boards and no independent benchmark had been published. The only independent test circulating at launch was a small comparison site running the same eight prompts on H3-Image and GPT Image 2.6, one try each — useful as a first look, not as evidence. One pricing aggregator's own test of six 4K jobs measured 23–39 seconds, with a median of about 28 — against MiniMax's under-10-seconds claim on its own platform. The gap may be the difference between MiniMax's infrastructure and a third-party reseller's; either way, "10 seconds" is a vendor number until someone reproduces it on the official API.
Why this launch matters
H3-Image lands in a month when the image-generation race has been running almost entirely on proprietary models — OpenAI's Images 2.6 arrived a week ago with GIFs and two new API models, and no weights at all. MiniMax's answer is the opposite bet: a big, fast, openly downloadable model that narrows the gap between open-weight and proprietary image quality, at half to two-thirds the price.
That bet only works if the license lets people use it. MiniMax is not the first Chinese lab to ship open weights with a territory-limited license — Baidu's ERNIE 5.5 Flash Base did the same thing a week earlier, excluding the EU, UK and South Korea — but H3-Image is the sharpest version of the pattern yet: a model good enough that creators in the excluded regions will genuinely want it, and a license that says no. The open-weight image race is now a licensing race too, and the fine print is part of the product.
For creators, the practical advice is simple. If you are outside the excluded territories and you work in image generation, H3-Image is worth a test run this week while the 50% launch discount holds — the 16-reference editing alone is a capability worth trying. If you are in the US, EU, UK or South Korea, read the license before you download anything, and treat the API as the only route that might be available to you. And wherever you are, remember that a launch-day speed claim is a vendor claim: the model is a day old, and nobody outside MiniMax has had the weights long enough to say what it really does.
Sources and classification
Classification: Official Announcement (MiniMax, 11 November 2026; the company's blog post is timestamped 10 November 2026, 16:23 UTC — 11 November, 00:23 in MiniMax's Beijing time zone — and the public announcement on X and in the press followed in the early hours of 11 November). Launch facts are from MiniMax's own announcement and model materials; speed, quality and pricing figures are MiniMax's own claims, attributed as such above, with no independent benchmark published yet. Checked 12 November 2026.
- MiniMax Blog, "MiniMax H3-Image: Bringing 4K Image Generation into the 10-Second Era" (10 November 2026, 16:23 UTC) — read first-hand: 4K (4096×4096) in under 10 seconds; up to 5× the speed of mainstream alternatives; up to 16 reference images; text-to-image, reference-based generation and image editing in one checkpoint; open weights; Standard mode 1.5–2× as fast as mainstream alternatives, Highspeed mode 3–4×; pricing 50–70% of comparable mainstream offerings; 50% off 4K-and-above generations in launch week; availability via MiniMax API Platform, MiniMax Code and Hailuo AI; Hugging Face repository MiniMaxAI/MiniMax-H3-Image: https://www.minimax.io/blog/minimax-h3-image
- MiniMax H3-Image model card and GitHub repository — 33.1B-parameter diffusion transformer distilled to six sampling steps; 68.65 GB BF16 checkpoint; deployment via Diffusers, SGLang, vLLM-Omni and ComfyUI; Community License Agreement excluding the US, EU, UK and South Korea; written authorisation required above $1M annualised revenue or for hosted services; model name must be displayed in commercial interfaces; outputs may not train other models; no masked inpainting, transparent backgrounds, user-specified seeds or preview images (per the GitHub README): https://huggingface.co/MiniMaxAI/MiniMax-H3-Image
- KOCPC, "MiniMax releases H3-Image, its strongest open-source image generation model" (11 November 2026) — announcement on X in the early hours of 11 November; weights uploaded to Hugging Face 10 November, 14:50 UTC; 50% off 16MP+ generations in the first week; license terms corroborated: https://en.kocpc.com.tw/archives/27381
- ToolScout, "MiniMax H3-Image: 4K Images in 10 Seconds, With One Big Catch" (11 November 2026) — launch-week discount window to 17 November, 16:00 UTC, for images of 16.6 megapixels and above; no Artificial Analysis image-board entry as of 11 November; Pixel Dojo eight-prompt comparison vs GPT Image 2.6: https://toolscout-e2i.pages.dev/posts/minimax-h3-image-4k-open-weights-license-catch
- Yir.ai, "MiniMax H3 Image API" (11 November 2026) — independent 4K test timings (23–39 seconds, median ~28, on KIE) vs MiniMax's under-10-seconds claim; single 16 GB RTX 5080 rendering report: https://yir.ai/models/minimax-h3-image