AI CURATED HOT STORY · 11/10/2026
MiniMax has officially open-sourced its H3-Image image generation model, releasing the model weights. The model supports text-to-image, image reference, and image editing, and can generate 4096×4096 (about 16MP) 4K images in 10 seconds, which the company claims is about 5 times faster than mainstream levels. It accepts up to 16 reference images and boasts the strongest instruction-following capability among open-source models.
According to multiple sources, H3-Image is derived from the H3 base and is available via MiniMax API, MiniMax Code, and Hailuo AI, and has entered Arena's battle mode for blind user voting. On the same day, Moore Threads announced that its MTT S5000 accelerator and MUSA software stack were adapted, enabling single-card generation of 4096×4096 images. The adaptation reuses components like SGLang, MATE, and MCCL, focusing on optimizing the high overhead of diffusion model attention in multi-reference scenarios.
The open-source release signals a further shift of high-quality image generation toward open ecosystems. The speed of generating 4K images in 10 seconds and multi-reference support could lower barriers for professional creation, while the Day-0 adaptation by a domestic accelerator highlights accelerating hardware-software co-design, paving the way for Day-0 support of future image and video generation models.
The line compares only sources observed throughout; gap hours are interpolated to keep the trend continuous.