Skip to main content
AI RADARModels
0•11/11/2026, 17:19•2 min read

MiniMax releases H3-Image with native 4K

Executive TL;DR30s read

The new image generator built from the H3 video model produces native 4096×4096 images in six steps and supports long prompts and multiple references.

MiniMax releases H3-Image with native 4K

From the H3 video model, MiniMax has created an image generator. It renders native 4096×4096 images in six steps. You can provide prompts up to 32,000 characters and up to 16 reference images.

Text-to-image generation at 4K took about 10.3 seconds in Highspeed mode and 16 seconds in Standard mode. In the same test, Nano Banana 2.1 with thinking medium took 50 seconds. In the API, such generation in Standard costs about $0.07, with a 50% launch discount currently available. Until November 23 inclusive, fal offers 3–5 free generations.

H3-Image is a diffusion transformer with 33.1B parameters. Together with the encoder and VAE, it requires about 135 GB to download. For 24 GB GPUs, SGLang suggests offloading layers to system RAM. ComfyUI already has templates. Enthusiasts have also released a two-step LoRA, but its quality still needs to be verified.

Announcement Weights Code Try on fal

Why it matters

Analysis

This release shows that open AI image models are closing the gap with commercial services in speed, quality, and inference cost.

Discuss in community

Share your questions and insights with developers

Join Telegram
+5 Points
← Back to news feed

Related in Ecosystem

Practical guides, core terminology, and agent skills