China's MiniMax Gives Away an Image Model It Says Makes 16-Megapixel 4K Pictures in Under 10 Seconds

AI image generators usually force a tradeoff: fast and fuzzy, or sharp and slow. Chinese AI startup MiniMax says its new model doesn't.

Late on Nov. 10, Beijing time, MiniMax announced MiniMax H3-Image and released its model weights. The model can create pictures from text, generate new images based on reference photos, and edit existing images.

The backstory

MiniMax is one of China's better-known AI startups and runs products such as the Hailuo AI video app. A few months ago it released a video model called H3, also with open weights. H3-Image is built on top of H3, with the audio components removed and the model tuned specifically for creating and editing still images.

What does "open weights" mean? A model's weights are essentially the file that holds everything the AI learned. Releasing them means anyone can download that file from sites like Hugging Face and run or modify the model on their own hardware, instead of paying per image through the company's service.

The numbers MiniMax is claiming

Speed: MiniMax says H3-Image can produce a 4096-by-4096-pixel image, about 16 megapixels, in under 10 seconds, and that it's up to five times faster than mainstream alternatives. For comparison, many AI image tools default to 1 to 4 megapixels. One caveat: the company's blog post doesn't say what hardware produced that 10-second figure, and running it on a home graphics card could be much slower.

References: It can take up to 16 reference images at once. An online seller, for example, could feed in several photos of one product and get ad images in different settings while keeping the product looking consistent.

Size: According to Chinese outlet Caijing, citing technical details on Hugging Face, H3-Image is a 33.1-billion-parameter diffusion transformer with 50 layers that uses Alibaba's Qwen3-VL as its text encoder. More parameters usually means finer detail, but also more graphics memory needed to run it.

On price, MiniMax says its API charges 50% to 70% of what comparable mainstream services do, with 4K-and-above generations half off during launch week.

What we know and don't know

Claims of near-photographic realism and best-in-class prompt following among open models come from MiniMax itself; there's no thorough independent evaluation yet. AI platforms including fal and Runware added the model on launch day, which shows industry interest but doesn't prove the quality claims.

Why it matters

High-resolution, high-quality AI image generation has mostly lived inside a handful of closed commercial products. A model that makes 4K images in seconds and can be downloaded freely lets small companies and independent developers build their own design and e-commerce tools without paying a big provider for every image. It's also another sign that Chinese AI companies are extending their open-model strategy from text into images and video.

What it means for you

If you run an online store, make social media content, or design flyers, models like this will likely show up inside the editing and design apps you already use, making images faster, sharper, and cheaper. The flip side: convincing fake photos get easier to make, too, so a little extra skepticism about images you see online is going to matter more.

Original sources published: MiniMax official X account (@MiniMax_AI) announcement, Nov. 10, 2026, 23:40 Beijing time (UTC+8); Caijing report, Nov. 11, 2026, 16:07 Beijing time