Felix Laurent
MiniMax released H3-Image on November 10, 2026, an image generation and editing model with downloadable weights. The company says it makes 4K images of about 16 million pixels in 10 seconds and accepts up to 16 reference images. It is available through the MiniMax API, MiniMax Code and Hailuo AI, with 50% off the largest images until November 17.
Two details matter more than the headline. MiniMax published no benchmark for the model, so its quality claims are still untested. And its license does not allow use of the weights in the United States, the European Union, the United Kingdom or South Korea.
What MiniMax released with H3-Image
H3-Image comes from the same family as MiniMax H3, the video model that we tested against Seedance 2.5 and Grok Imagine. It is that architecture with the audio layers removed. One checkpoint handles both text-to-image and editing, and it draws at full size without a separate upscaler.
| Spec | MiniMax H3-Image |
|---|---|
| Parameters | 33.1B diffusion transformer |
| Sampling steps | 6, distilled, no classifier-free guidance |
| Output | About 16 MP, from 4096×4096 to 6272×2688 |
| API size limit | Up to 20 million pixels, ratios from 1:3 to 3:1 |
| Reference images | 1 to 16 through the API |
| Images per request | 1 to 10 |
| Weights | 68.65 GB, plus a 66.7 GB text encoder |
| Not supported in the API | Masked edits, transparent backgrounds |
The 16 MP target is the main selling point. MiniMax notes that most image models today output 1 to 4 MP. It also claims "the strongest instruction following among open-source models", though the model card gives no test to support that.
The lack of masks and transparency is a real gap for design work. Seedream 5.5, released a day earlier, ships with native transparent output. The 10-second claim raises a separate question.
How fast H3-Image is
MiniMax has not said what hardware produces a 4K image in 10 seconds. The API also comes in a Highspeed version at about twice the output price, and the company publishes no timing for either tier.
The closest thing to a measurement comes from the vLLM project's deployment recipe, which covers images up to about 2K. These are its results on Nvidia H200 GPUs:
| Setup | Task | Time | Peak memory per GPU |
|---|---|---|---|
| 4× H200 | 2720×1536 image | About 0.8 s | – |
| 4× H200 | 2048×2048 edit, one reference | 1.35 s | – |
| 1× H200 | 2048×2048 image | 3.2 s | 73.4 GiB |
| 1× H200, model offloaded | 2048×2048 image | 8.3 s | 14.7 GiB |
A 16 MP image has four times the pixels of a 2048×2048 one, so it takes longer. Running the full model locally is data center work for now. Offloading cuts the GPU memory needed, but the download alone is about 135 GB.
Most people will use the API instead, and that is where H3-Image looks strongest.
H3-Image price compared with rival image models
MiniMax bills by tokens and publishes example prices per image. The standard model costs $0.02179 at 1K, $0.02333 at 2K and $0.06879 at 4K. The launch discount cuts the 4K price to $0.03439 until November 17 at 16:00 UTC.
| Model | Largest output | Price at about 1K | Price at largest size |
|---|---|---|---|
| MiniMax H3-Image | About 16–20 MP | $0.022 | $0.069, or $0.034 with the discount |
| Nano Banana 2.1 | 4K | $0.034 | $0.113 |
| GPT Image 2.6 Sunburst | About 8.3 MP | $0.21 at max quality | Not listed per image |
| Seedream 5.5 | About 8 MP, per WaveSpeed | From $0.02, per WaveSpeed | Not listed |
At 4K, H3-Image costs about 39% less than Nano Banana 2.1 at full price. Price says little about quality, however, and quality is the part nobody can compare yet. H3-Image was absent from both Artificial Analysis image arenas when we checked on November 11. GPT Image 2.6 Sunburst leads the text-to-image board there at 1259 Elo.
Cost also depends on where you are allowed to run the model.
What the H3-Image license allows
MiniMax calls the model open-weight, and the wording is accurate. The weights are public, but the MiniMax-H3-Image Community License is far narrower than an open source license. Its main terms:
- The license covers the world except the US, the EU, the UK and South Korea, and use there needs a separate deal with MiniMax
- Anyone with annual revenue above $1 million needs written authorization
- Anyone offering the model as a hosted service to others needs written authorization too
- Commercial products must display "MiniMax-H3-Image" in their interface
- Outputs cannot be used to improve any other AI model
The territorial rule also covers outputs. The license says the works and "any of their Outputs or results" may not be used outside the permitted area. A studio in Germany cannot simply download the weights and generate its campaign images locally. These terms apply to the weights, and the hosted API runs under MiniMax's own service terms.
MiniMax was also not first to turn H3 into an image model. On November 3, Pruna AI open-sourced P-H3-Image, built from the H3 video model. Pruna says it is seven times smaller than H3 and makes a 4 MP image in 2.6 seconds. With no benchmarks from MiniMax, early users are the only other evidence so far.
What early testers made with H3-Image
Creators with early access began posting results a day before the launch. The X user Arminn_Ai, shared a first set of four images from Hailuo. Kiyoshi Shin, a Japanese AI content developer, tested reference handling. Shin generated a three-view drawing of a mech, then used it as a reference to place the robot in a desert scene. The post says situational setups now work freely.
These are early-access posts tagged for the launch, so they lean positive. We found no independent speed test or side-by-side comparison yet. Community interest is visible on Hugging Face, where the model passed 300 likes in its first day.
Is MiniMax H3-Image worth trying?
For teams that need large images at a low price, yes, through the API and during the discount week. A 4K image for about three cents is cheap enough to test on real work. Developers in the US, the EU, the UK or South Korea should treat the weights as unavailable until MiniMax grants them a license. The same goes for any company above $1 million in revenue. The bigger open question is quality, since a 16 MP image is only useful if the extra pixels hold real detail. Until H3-Image appears on the arenas, the 10-second speed and the instruction-following claim remain MiniMax's word.
Sources
- Meet MiniMax H3-Image, MiniMax on X – November 10, 2026
- MiniMax-H3-Image model card, Hugging Face
- MiniMax-H3-Image Community License Agreement – November 10, 2026
- Image generation guide, MiniMax API docs
- Pricing overview, MiniMax API docs
- MiniMax-H3-Image recipe, vLLM
- Artificial Analysis text-to-image leaderboard – accessed November 11, 2026
- Artificial Analysis image editing leaderboard – accessed November 11, 2026
- P-H3-Image release, Pruna AI on X – November 3, 2026
- MiniMax releases open-weight H3-Image model, Crypto Briefing – November 10, 2026
