MiniMax H3-Image: 4K AI Images in Under 10 Seconds, With a Catch

MiniMax released H3-Image on November 10, 2026, and the headline is hard to ignore: a 4K image in under 10 seconds, from a model whose weights anyone can download. The company says it is up to five times faster than mainstream alternatives and priced at 50 to 70 percent of what comparable services charge. Both claims come from MiniMax itself, and both need the fine print read before you plan anything around them.

The fine print is where this model gets interesting. The per-image prices check out against MiniMax's own pricing page, and at 4K the saving over Google's Nano Banana 2.1 is real. But the license that ships with the weights excludes the United States, the United Kingdom, the European Union and South Korea from local use, which covers most of the audience this site writes for. This review walks through what H3-Image does, what it costs, and where the catches sit.

What MiniMax H3-Image is

H3-Image is an open-weight image generation and editing model built on MiniMax's H3 family, released on November 10, 2026. One checkpoint handles text-to-image generation and image editing, with up to 16 reference images for edits, addressed by position in the prompt. It is available as downloadable weights on Hugging Face, through the MiniMax API Platform, in MiniMax Code, and in the Hailuo AI web app.

Do not confuse it with MiniMax's M3.1 Flash, the language model covered in this site's October speed comparison. H3-Image is a separate product from the H3 family, built for pictures rather than text. The model is a 33.1-billion-parameter dense diffusion transformer, distilled to six sampling steps with no classifier-free guidance. In practice that means every generation runs the same short, fixed sampling schedule, which is part of why MiniMax can promise the speed it does.

The weights are a large download. The diffusion transformer and VAE total 68.65 GB in BF16, and the text encoder adds another 66.7 GB, pulled from the separate MiniMax-H3 repository. A first local install is roughly 135 GB before you generate a single image. That number matters more than it looks, as the section on self-hosting below explains.

4K in under 10 seconds: what the speed claim means

MiniMax's headline promise is specific: a full-resolution 4K image, 4096 by 4096 pixels, in under 10 seconds. The company also says the Standard mode runs 1.5 to 2 times faster than mainstream alternatives, while a Highspeed mode reaches 3 to 4 times. All of those figures are MiniMax's own, with no independent timing published yet, one day after launch.

The resolution story is the more concrete part. H3-Image generates about 16 megapixels natively at every aspect ratio, with no upscaler involved. The model card lists 4096 by 4096 at 1:1, 4736 by 3552 at 4:3, 5024 by 3360 at 3:2, 5472 by 3072 at 16:9, and 6272 by 2688 at 21:9. Through the API the output area runs from about 4 megapixels to a 20-megapixel cap, with 4K as the default area level. The card says results are generally best at around 16 megapixels.

Two API limits are worth knowing before you build on this. Prompts can be up to 32,000 characters, but with prompt enhancement switched off only about the first 4,096 tokens take effect. And the text-to-image endpoint returns at most 10 images per request. MiniMax ships a prompt expander to work around the prompt limit: a hosted Context-IR service, or a local 17.9 GB model called H3-Image-PE. There is no negative-prompt support, and the API rejects any guidance scale other than 1.

What it costs

MiniMax prices H3-Image by the token, like a language model. The Standard tier charges $2.30 per million input text tokens, $1.50 per million input image tokens, and $3.80 per million output image tokens. Highspeed keeps the same text and image input rates but doubles the output rate to $7.60 per million tokens, with input image tokens at $3.00. Those are the rates on MiniMax's own pricing documentation, read on November 10.

Token rates are hard to compare, so MiniMax also publishes worked examples per image. These are examples, not fixed prices: the pricing page warns that token usage varies, that some 1K and 2K example amounts were adjusted for presentation, and that actual charges are based on token usage. With that caveat, the Standard-tier examples are:

Highspeed costs more at every size. A 4K text-to-image generation is $0.13001 on Highspeed, against $0.06879 on Standard. A 4K edit with three reference images is $0.15407 on Highspeed, against $0.07914 on Standard. Speed is a paid upgrade here, not a free tier.

Launch week carries a 50 percent discount, but it is narrower than the headline suggests. It applies only when the delivered image is at least 16,600,000 pixels, which means 4K and above, and it expires on November 17, 2026 at 16:00 UTC. A discounted 4K text-to-image generation costs $0.03439 on Standard. Anyone pricing a pipeline on that number after November 17 is pricing on a promotion that no longer exists.

In pounds and Canadian dollars, at November 10, 2026 market rates of 1 US dollar to 0.745 British pounds and 1.411 Canadian dollars, a Standard 1K image is about 1.6 pence or 3.1 Canadian cents, and a Standard 4K image is about 5.1 pence or 9.7 Canadian cents. Those are approximations, and the dollar figure is what MiniMax actually bills.

The monthly math is where the saving becomes tangible. At list prices, 20,000 images a month at 2K costs $466.60 through MiniMax's API, or about 347.48 pounds and 658.50 Canadian dollars. The same volume through Google's Nano Banana 2.1 costs $1,008, a gap of $541.40 a month. At 4K, 2,000 images a month costs $137.58 on MiniMax against $226 on Google, a gap of $88.42. Only successful images are charged, according to MiniMax's API documentation, so failed generations do not add to the bill.

The license catch

Read the license before downloading the weights, because it is not the license most people assume when they hear open weights. H3-Image ships under the MiniMax H3-Image Community License, dated November 10, 2026. It is a custom agreement, not MIT or Apache, and its first page limits the grant to an Applicable Territory defined as worldwide excluding the European Union, the United Kingdom, the Republic of Korea and the United States of America.

The text is blunt: use outside the Applicable Territory is not authorized by the agreement. MiniMax says it will keep evaluating the laws and compliance requirements for the excluded territories, and invites people there to contact it about obtaining a license. For readers in the United States or the United Kingdom, that means the weights are not licensed for local use as published. The API, MiniMax Code and Hailuo AI are separate services with their own terms, and MiniMax has not said publicly whether API access is region-restricted in the same way, so do not assume the API inherits either the permission or the ban.

Two more conditions apply even inside the permitted territory. If your annualized revenue, counting affiliates, exceeds 1 million US dollars, or if you offer hosted services to third parties, you need separate prior written authorization from MiniMax. And any commercial interface that uses the model must display the words MiniMax-H3-Image prominently. Outputs also may not be used to train or improve other models. None of this makes the model unusable, but it does make the word free misleading for a large share of commercial users.

H3-Image vs Nano Banana 2.1

Google's Nano Banana 2.1 is the natural comparison, because it is the measured model at the same price point. It has been generally available since October 6, 2026 as gemini-nano-banana-2.1, and Google's own pricing page lists $30 per million output image tokens, which works out to $0.0336 per 1K image, $0.0504 per 2K image, and $0.113 per 4K image. A batch tier halves those output rates for jobs that can wait.

On price, MiniMax wins at every resolution on the Standard tier, by about 35 percent at 1K and about 39 percent at 4K. The Highspeed tier is a different story: at $0.13001 per 4K image it costs more than Google's $0.113, so the speed upgrade erases the price advantage entirely. The image-model race is moving fast: ByteDance's Seedream 5.5, reviewed here, arrived a day earlier with a different pitch, 100 editing rounds and transparent PNGs, while fal's FLUX 3 Max pushes AI video rather than stills.

On evidence, Google wins by default. Artificial Analysis scores Nano Banana 2.1 at Elo 1160 on its text-to-image board and 1137 on its image-editing board, fourth place on both, over roughly 6,000 samples each. H3-Image appears on neither board, in any position, with any sample count. MiniMax's model card calls it a leading open-weight image model and claims the strongest instruction following among open-weight models, but the card publishes no evaluation table, no GenEval or DPG-Bench numbers, and no human preference scores. The only quality evidence in public is the showcase grid in the repository, which is the kind of evidence designed to look good.

The two models also differ in what they hand you beyond the image. Google's 14 reference images are typed, 10 for objects and 4 for characters, which tells the model what role each image plays. MiniMax gives you 16 untyped slots, more headroom and less structure. Google grounds generation in web and image search and marks every output with SynthID. MiniMax states no watermarking scheme anywhere in the repository or the license. If your deliverable needs provenance marks, that gap decides the choice before price does.

Running it on your own hardware

Self-hosting is where the open weights earn their keep, and where the costs hide. The model runs through SGLang, vLLM-Omni, diffusers and ComfyUI, and MiniMax's serving guidance names NVIDIA H200, RTX PRO 6000 Blackwell and AMD Instinct MI350X hardware, with multi-GPU sequence parallelism as the documented way to get acceptable latency at full resolution. This is datacenter-class hardware, not a laptop project.

Do the download math first. The checkpoint is 68.65 GB and the text encoder is another 66.7 GB, so a first install is about 135 GB across two repositories. At MiniMax's API prices, a 1K image costs about two cents, which buys a very large number of generations before self-hosting pays for itself in hardware and electricity. The weights are valuable for control, for offline work, and for the license terms, not for unit economics at small scale. If neither the download nor the API bill appeals, the free options in our roundup of free AI image generators still cover the basics.

One practical detail: the local prompt expander, H3-Image-PE, is a separate 17.9 GB download in NVFP4. Without prompt enhancement, the model reads only about the first 4,096 tokens of a prompt, so long, detailed prompts need the expander or the hosted Context-IR service to take full effect. The expander's editing side is described as prompt-only and still in development, so it currently helps text-to-image more than editing.

What is still unproven

H3-Image is one day old in public, and the honest summary is that the cheaper model is cheaper and the measured model is measured. Nothing in the launch material closes the gap between those two facts. The speed figures, the prompt-adherence claim and the quality claims are all MiniMax's own, with no independent test, no benchmark table and no Artificial Analysis placement as of November 10.

Three gaps deserve specific attention. First, there is no published evaluation of any kind in the model card, so claims about instruction following cannot be checked against a competitor on the same test. Second, there is no stated watermarking or provenance scheme, which matters for anyone publishing generated images commercially. Third, the hosted prompt expander that makes short prompts work well is only half finished on the editing side, according to the repository's own description.

None of this makes H3-Image a bad model. It makes it an unmeasured one. The sensible pattern for now is to route production work to the model with a scoreboard and send a slice of traffic to H3-Image with a fixed prompt set, so you build your own evidence before deciding. A month from now, when independent scores exist, the price gap stops being a proxy argument and becomes an actual trade.

Common mistakes

Treating the launch discount as the price is the first mistake. The 50 percent cut applies only to images of at least 16.6 megapixels and ends November 17, 2026 at 16:00 UTC. A 1K or 2K image never qualified, and a 4K image stops qualifying next week.

Assuming the weights are free to use anywhere is the second. The community license excludes the United States, the United Kingdom, the European Union and South Korea, and requires written authorization above 1 million dollars in annualized revenue or for hosted services. Downloading the weights is not the same as being licensed to run them.

Skipping the prompt expander is the third. A long prompt sent with enhancement off is truncated at about 4,096 tokens, so the detail you wrote past that point has no effect. Either keep prompts short, use the hosted Context-IR service, or install H3-Image-PE locally.

Finally, do not price a project on the Highspeed tier by accident. At 4K it costs more per image than Nano Banana 2.1, which removes the entire price argument for choosing MiniMax. Use Standard unless you have measured that the speed difference is worth the premium.

Is MiniMax H3-Image free?

The weights are free to download, and the API is pay-as-you-go with no subscription. But the community license restricts where the weights may be used and requires authorization above 1 million dollars in annualized revenue or for hosted services, so free to download is not the same as free to use commercially.

Can I use MiniMax H3-Image in the US or UK?

Not under the community license as published. The license excludes the United States, the United Kingdom, the European Union and South Korea from the territory where use is authorized. MiniMax invites people in those territories to contact it about a separate license. Whether the hosted API carries the same restriction has not been stated publicly.

How much does MiniMax H3-Image cost per image?

On the Standard tier, a text-to-image generation costs $0.02179 at 1K, $0.02333 at 2K, and $0.06879 at 4K, with a 4K image discounted to $0.03439 until November 17, 2026. Edits cost a little more and rise with the number of reference images. Actual charges depend on token usage, so these are MiniMax's worked examples rather than fixed prices.

Is H3-Image better than Nano Banana 2.1?

It is cheaper at list prices on the Standard tier, by roughly a third at 1K and 4K, and it generates larger images, up to 20 megapixels against Google's 4K ceiling. But Nano Banana 2.1 is the only one of the two with independent scores, ranking fourth on both Artificial Analysis image boards. Until H3-Image is measured the same way, cheaper is the only claim that holds up.

What hardware do I need to run it locally?

About 135 GB of downloads, a 68.65 GB checkpoint plus a 66.7 GB text encoder, and datacenter-class GPUs. MiniMax's serving guidance names the NVIDIA H200, RTX PRO 6000 Blackwell and AMD Instinct MI350X, with multi-GPU parallelism for full-resolution work. For most users the API is cheaper than the hardware.

Does H3-Image watermark its outputs?

No watermarking scheme is stated anywhere in the model card, the repository or the license. Google marks every Nano Banana 2.1 output with SynthID. If your workflow needs provenance marks on generated images, H3-Image does not offer one as published.

H3-Image is an interesting release: real 4K resolution without an upscaler, a six-step sampler that makes speed a feature rather than a trade-off, and per-image prices that undercut Google by a third on the Standard tier. The license exclusion covering the United States, the United Kingdom, the European Union and South Korea is the fact that changes the decision for most readers, and the absence of any independent measurement is the fact that should slow down everyone else. Try it through the API, where the cost of finding out is a few cents an image, and keep the production work on the model with a scoreboard until H3-Image earns one.

Sources

• MiniMax, "MiniMax H3-Image: Bringing 4K Image Generation into the 10-Second Era," official blog, November 10, 2026
• MiniMax API documentation, pricing (Standard and Highspeed token rates, per-image example costs, 50% launch discount terms)
• MiniMax API documentation, text-to-image endpoint (models, prompt limits, image counts, output formats)
• Hugging Face model card, MiniMaxAI/MiniMax-H3-Image (architecture, resolutions, deployment, prompt expander)
• MiniMax H3-Image Community License Agreement, November 10, 2026 (excluded territories, authorization threshold, branding requirement)
• Google, Gemini Developer API pricing (Nano Banana 2.1 rates: $0.0336 per 1K, $0.0504 per 2K, $0.113 per 4K image)
• OrcaRouter, "MiniMax-H3-Image vs Nano Banana 2.1: price vs proof," November 10, 2026 (Artificial Analysis scores, launch timing, monthly cost examples)