AMD GPUs Support MiniMax-H3 Model at Launch

AMD has introduced Day 0 help for MiniMax-H3, the newest video technology mannequin from Chinese language AI agency MiniMax, on its Intuition MI355X and MI300X GPUs. This collaboration permits the highly effective multimodal mannequin to run effectively on AMD’s {hardware}, marking a big milestone for AI-driven video and audio technology.

MiniMax-H3, launched on July 31, 2026, represents a generational leap in video synthesis. In contrast to its predecessor, the Hailuo sequence, which separated duties like text-to-video (T2V) and image-to-video (I2V) into distinct pipelines, H3 unifies textual content, picture, video, and audio inputs right into a single framework. The mannequin generates as much as 15-second clips at 2K decision (1440p) with synchronized stereo audio, surpassing the 1080p, 10-second output of earlier fashions. It additionally introduces capabilities like instruction-based video modifying and reference-driven animation, making it one of the vital versatile open-weight video fashions obtainable.

AMD’s help depends on its ROCm platform and SGLang framework to optimize MiniMax-H3’s compute-heavy structure. By using 8-way sequence parallelism and leveraging AMD’s AI Tensor Engine (AITER) kernels, the GPUs guarantee easy processing of the joint video-audio latents that type the core of H3’s output. That is crucial for sustaining alignment between movement and sound, a trademark of H3’s capabilities.

The mannequin structure itself is extremely modular. MiniMax-H3 splits duties throughout two partitions: FL2VA for text-to-video and frame-conditioned enter, and Ref2VA for reference-driven technology. Every partition consists of parts like a multimodal textual content encoder, specialised video and audio VAEs, and a diffusion transformer that denoises latents in tandem. This design not solely enhances efficiency but in addition makes deployment scalable throughout totally different {hardware} setups.

Past technical developments, MiniMax-H3’s open-weight launch underneath the MiniMax Neighborhood License is a serious draw for builders. It lowers boundaries to entry for these experimenting with video technology, positioning H3 as a dominant participant within the open AI panorama. The mannequin is already obtainable on platforms like Atlas, the place customers can entry it for $0.14 per second at 2K decision.

For AMD, this launch highlights the rising synergy between AI workloads and its Intuition GPU line. By enabling environment friendly deployment of a cutting-edge mannequin like MiniMax-H3, the corporate alerts its competitiveness within the AI {hardware} market, the place NVIDIA’s dominance has lengthy been unchallenged. The collaboration additionally underscores the rising demand for high-performance GPUs able to dealing with subtle diffusion-based fashions.

Trying forward, AMD’s help for MiniMax-H3 units the stage for additional optimizations, together with batch processing capabilities and improved kernel efficiency. As AI-driven video technology good points traction, each AMD and MiniMax seem well-positioned to capitalize on the rising curiosity in multimodal content material creation.

Picture supply: Shutterstock