MiniMax releases H3-Image: a 33B open-weight image model that renders 16MP in six steps

MiniMax has shipped its own image model. H3-Image, released on November 10 with open weights on Hugging Face, turns the company's H3 video-model family into a dedicated image generator and editor — one that renders 16-megapixel images in six sampling steps and takes up to 16 reference images in a single edit.

The release landed a day after the model surfaced as an early-access preview inside MiniMax's Hailuo app, and its rollout was slightly ragged: the announcement post went out a little before the weights themselves, and the Hugging Face link briefly returned a 404 before the repository went live later in the morning. By late morning, the card, weights and licence were all live.

What shipped#

H3-Image is a single checkpoint that does both text-to-image generation and image editing. The diffusion transformer has 33.1 billion parameters and follows the H3 base design with the audio branch removed; prompts and reference images are encoded by the H3 text encoder, and the image VAE decodes the result. The headline numbers, from MiniMax's own model card:

The pitch is speed at high resolution. Most open image models top out at 1–4 megapixels; H3-Image renders four times that directly, without an upscaling pass, and does it in six steps. MiniMax also claims the strongest instruction following among open-source models — its own claim, not yet independently tested — and is offering 50% off 16-megapixel-and-up generations during launch week.

Six steps, 16 megapixels: why the numbers matter#

Six sampling steps is the number to watch. A model distilled down to a handful of steps can generate a 16-megapixel image fast enough for interactive use — MiniMax advertises 4K-by-4K images in 10 to 20 seconds — which is what makes the resolution usable in practice. The trade-off with heavy distillation is usually detail and prompt adherence, so the real test is whether output holds up at full size, not just in downscaled showcase grids.

The 16-reference-image ceiling matters for a different reason. Multi-reference editing — blending a character, a product, an outfit and a location into one frame — is where image models are currently competing hardest, and most rivals cap out well below 16. If H3-Image keeps characters consistent across that many inputs, it becomes a serious tool for storyboarding, advertising mockups and product shots — but that is unproven outside MiniMax's own examples.

Not Pruna's model — don't confuse the two#

Anyone who followed H3 news last week may feel a sense of déjà vu. On November 3, Pruna AI released P-H3-Image, a third-party distillation that rebuilt the H3 video model into an image generator: about 20 GB, 8 steps, up to 8 reference images. It was a clever piece of engineering, but it was not MiniMax's model.

H3-Image is the first-party version — full-size, 33.1 billion parameters, 68.65 GB for the transformer alone plus a separate 66.7 GB text encoder — and it differs in every dimension that matters: bigger, faster per step, higher resolution, twice the reference images. The two share a licence family and a name fragment, which is why they will be confused. Check the repository before downloading: MiniMaxAI/MiniMax-H3-Image versus PrunaAI/P-H3-Image.

The licence is the catch#

Open weights, in MiniMax's case, comes with a map. The MiniMax-H3-Image Community License — dated November 10, 2026, the release day — grants use worldwide except the European Union, the United Kingdom, South Korea and the United States. Anyone in those territories needs a separate licence from MiniMax. On top of that, any user whose annual revenue tops $1 million, or who offers the model as a hosted service, needs separate written authorization, and commercial products are expected to display "Powered by MiniMax-H3-Image". MiniMax claims no rights over images people generate with the model.

That is the same territorial carve-out the H3 video model shipped with, and it shapes who this release is actually for: researchers and builders outside the excluded regions, plus anyone willing to go through MiniMax's licensing channel. For a North American or European reader, the honest summary is that the weights are downloadable and the API is usable, but the licence does not cover you.

What to watch#

Three things will decide whether H3-Image matters beyond launch day. First, independent comparisons: how its 16-megapixel output and 16-reference editing actually stack up against GPT Image, Nano Banana and Seedream once people run the same prompts through all of them. Second, the tooling: ComfyUI templates and Diffusers support are promised, and a 135 GB total download (transformer plus text encoder) means most users will wait for quantized versions. Third, the licence question — whether MiniMax opens the excluded territories, as the community will inevitably ask, or holds the line. The model is real and the weights are live; the rest is now up to the people who run it.