Moore Threads announces Day-0 support for the MiniMax H3-Image generation model
Moore Threads announced on November 11 that its MTT S5000 AI training and inference card with the MUSA software stack adapted the MiniMax H3-Image generation model on the day it was open-sourced.
Moore Threads announced on November 11 that its MTT S5000 AI training and inference card and MUSA software stack had completed adaptation of the MiniMax H3-Image image generation model on the day it was open-sourced, November 10. H3-Image is derived from H3 and offers text-to-image generation, image reference and image editing capabilities, producing a 4096 by 4096 pixel, roughly 16MP high-resolution image within 10 seconds at up to five times the mainstream market speed, along with solid instruction following and multi-reference understanding.
For H3-Image's core characteristics, Moore Threads performed full-chain adaptation and deep optimization on the MTT S5000. Diffusion-based image generation and editing models carry heavy attention overhead with a high attention share, and key-value expansion is especially pronounced in multi-reference scenarios, so overhead rises faster the more reference images are used. With its high compute density, the MTT S5000 can absorb the heavy compute load during understanding and high-resolution generation, releasing H3-Image's capabilities.
Moore Threads had previously completed attention and general matrix multiplication kernels along with communication and runtime infrastructure on the MiniMax H3 model. For H3-Image, the team reused mature components including SGLang, MATE and MCCL, drawing on visual generation model optimization experience, and established a foundation for Day-0 adaptation of further image and video generation models. Whether domestic compute platforms can ship usable support on the day a model is open-sourced directly determines the practical usability boundary of such open-weight models for developers in China.