Moore Threads MTT S5000 Completes Day-0 Adaptation for MiniMax H3-Image Image Generation Model
Moore Threads announced that its MTT S5000 AI training-and-inference integrated intelligent computing card and MUSA software stack completed adaptation on the same day MiniMax H3-Image was open-sourced. H3-Image is derived from H3 and supports text-to-image, image reference, and image editing. It can generate a high-resolution image of 4096 × 4096 pixels, approximately 16MP, in under 10 seconds, at a speed up to 5 times that of mainstream levels on the market. Moore Threads performed full-link adaptation and deep optimization based on the MTT S5000, reusing components such as SGLang, MATE, and MCCL.
Body · AI translation
Day-0 support: Moore Threads completes adaptation of the MiniMax H3-Image image generation model
2026-11-11 09:56 IT Home - Luyuan
Thanks to IT Home user Autumn_Dream for the tip-off!
IT Home, November 11 news: Moore Threads announced today that its AI training-and-inference integrated intelligent computing card MTT S5000 and MUSA software stack were adapted to the MiniMax image generation model H3-Image yesterday (the day it was open-sourced).
▲ Generation results of MiniMax H3-Image on a single MTT S5000
IT Home has learned that MiniMax H3-Image is an image generation model derived from H3, featuring text-to-image generation, image reference, and image editing capabilities. It can generate a 4096 × 4096 pixel, approximately 16MP (4K) high-resolution image within 10 seconds, at a speed up to 5 times that of mainstream levels on the market, while presenting rich, realistic details and possessing excellent instruction-following and multi-reference-image understanding capabilities.
In response to the core capability characteristics of MiniMax H3-Image, Moore Threads carried out full-link adaptation and deep optimization based on the MTT S5000:
Diffusion-based image generation and editing models have a relatively large attention mechanism overhead, with a high Attention proportion; especially in scenarios with multiple reference image inputs, Key / Value expansion is pronounced, and the more reference images there are, the faster the overhead rises. With its high computing power density, the MTT S5000 can readily handle the high computational load of such models during understanding and high-resolution generation, fully unleashing the performance of MiniMax H3-Image.
Moore Threads previously completed the construction of core operators such as Attention and GEMM, as well as underlying infrastructure such as communication and runtime, on the MiniMax H3 model. This time, for MiniMax H3-Image, the team combined its experience in optimizing visual generation models and reused mature component capabilities such as SGLang, MATE, and MCCL to complete the adaptation, while also laying the foundation for Day-0 adaptation of more image generation and video generation models in the future.