Overview
- MiniMax published H3-Image as an open‑source release with model weights, Hugging Face hosting, API endpoints and code links so developers can access and run the model.
- The company reports H3-Image can produce 4096×4096 (≈16MP) images in about 10 seconds and claims generation speeds up to five times faster than mainstream options, with these figures presented as vendor reports.
- H3-Image is described as supporting text‑to‑image prompts, image editing and multi‑reference conditioning with support for up to 16 reference images and enhanced instruction‑following ability, according to MiniMax.
- Moore Threads announced Day‑0 adaptation of H3-Image on its MTT S5000 card and MUSA stack and said it optimized core operators such as Attention and GEMM plus runtime components to reduce the heavy key/value expansion that comes with many reference images.
- MiniMax is offering a 50% discount for 16MP+ generation requests during the first week, a move that, together with immediate hardware optimizations, could accelerate real‑world testing and on‑prem adoption even though independent benchmarks are not yet available.