MiniMax H3-Image Renders Native 4K in 16 Seconds and Opens Its Weights

For years, most image generators lived with one compromise: the model draws at a lower resolution, and then an upscaler fills in "plausible" details. That is enough for previews. It is not enough for print, banners, and large layouts. MiniMax offers a different path: H3-Image generates 4096×4096 directly, with no separate upscaling step.

What this model is

H3-Image grew out of the H3 video model, which MiniMax reworked for still frames. Under the hood is a diffusion transformer with 33.1 billion parameters. It produces an image in six steps, which is very few for a model of this size.

Two more parameters matter in practice. The prompt can be up to 32,000 characters long, which means a whole brief fits into it: a brand description, the frame layout, style requirements. Up to 16 references can be passed in, for example a character, a product, a palette, and composition examples. For product teams this is closer to a working tool than a toy: a brand guide can be put straight into the request.

Speed and price

According to the announcement, text-to-4K takes about 10.3 seconds in Highspeed and 16 seconds in Standard. Nano Banana 2.1 with thinking medium took 50 seconds in the same test. The comparison should be read with a caveat. Thinking mode adds reasoning time before drawing, so the difference cannot be attributed entirely to the speed of generation itself. The post does not say what hardware was used for measurements, so the numbers should be rechecked on your own tasks.

Through the API, a 4K image in Standard costs about $0.07, and a 50% launch discount is currently in effect. Through November 23 inclusive, fal offers 3–5 free generations. That is enough to quickly run your own prompts without deploying anything yourself.

Running locally: heavy, but possible

The weights are open, and the code is on GitHub. But together with the encoder and VAE, you need to download about 135 GB. For 24 GB graphics cards, SGLang offers to offload some layers to system RAM. It works, but speed will drop noticeably, and the figures from the announcement do not apply to this configuration. ComfyUI already has ready-made templates, so you will not have to build the pipeline by hand.

The community has already released a two-step LoRA for acceleration. How much it degrades quality is not yet clear, and this needs to be checked separately.

What to do with this

If you are building a product with image generation, the cheapest step right now is this: take 10–20 of your real prompts, including long ones and ones with references, and run them through the free generations on fal. Look at what usually breaks in applied tasks: fine details, text in the frame, keeping a character consistent across references. Native 4K only makes sense if the gain in detail is visible without upscaling.

For local deployment, calculate disk and memory in advance. 135 GB, offloading, and unknown speed on a consumer card can eat up all the savings over the API. It is also worth reading the license on Hugging Face before integrating into a commercial product.

The overall shift is clear: 4K resolution and a long prompt context are becoming the baseline, and six-step speed makes this kind of generation suitable for interactive scenarios. Independent comparisons will show the rest.