MiniMax just did something the big image labs won't: it released a 4K image model you can actually download. H3-Image, launched on November 10, generates native 16-megapixel images in about 10 seconds — and the weights are sitting on Hugging Face right now, free for anyone to grab. If you've ever wanted a serious image model running on your own machine instead of behind someone else's API, this is the closest anyone has come.
What H3-Image actually does
The headline number is resolution without an upscaler in the middle. H3-Image generates natively at around 16 megapixels at any aspect ratio — 4096×4096 square, 5472×3072 at 16:9, even 6272×2688 ultrawide — in roughly 10 to 20 seconds, according to MiniMax. Under the hood it's a 33.1-billion-parameter dense diffusion transformer, distilled down to just six sampling steps. It also handles image editing and reference-based generation: you can feed it up to 16 reference images in a single edit, so keeping a character, product or style consistent across a set of images gets much easier. It's available three ways — through the MiniMax API, MiniMax Code, and Hailuo AI — and ComfyUI shipped native support on launch day, so local-generation fans didn't have to wait.
Open weights — with a big catch
Here's where the story gets interesting. "Open weights" sounds like "free for everyone," but read the fine print first. H3-Image ships under MiniMax's own Community License, and the model card excludes self-hosting in the European Union, the United Kingdom, the United States and South Korea. Some higher-revenue and hosted-service uses also need written authorisation. In plain English: you can download the 68.65 GB of weights, but whether you can legally run them where you live is a separate question. Hosted access through platforms like Venice sidesteps the restriction, but the "open" label comes with real asterisks.
What it costs, and what we don't know yet
MiniMax is pricing aggressively: the company says its rates run at 50 to 70 percent of comparable mainstream offerings, and launch week brings 50 percent off generations at 4K and above, until November 17. One hosted platform lists it at 4 cents per image at 1K and 2K, and 10 cents at 4K. But be honest about the gaps: nobody outside MiniMax has had the weights for more than a day, and as of November 11 there are no independent benchmarks — no entry on the major image-model leaderboards yet. Every speed and quality claim so far is MiniMax's own. The early hands-on tests look promising, but treat launch-day demos as demos.
Why it matters
This is MiniMax's second quiet launch in six weeks — its
M3.1-Flash-Preview coding model slipped out with zero fanfare in September — and the pattern is clear: ship fast, ship open, let the community do the marketing. While OpenAI's
GPT Image 2.6 and Google's image tools stay locked behind APIs, an open-weight 4K model changes who gets to build with top-tier image generation. If the quality holds up under independent testing, H3-Image could become the default image engine inside open-source tools — and the license catch is the one thing standing between "download it" and "use it anywhere."
4K in 10 seconds, open weights, 16 reference images — on paper, H3-Image is the most generous image model launch of the year. Just check the license before you download 68 GB of it.
Comments
Post a Comment