Darius Baruo
Aug 02, 2026 17:02
AMD Intuition GPUs allow Day 0 deployment of MiniMax-H3, a cutting-edge video era mannequin producing 2K clips with stereo audio.
AMD has introduced Day 0 assist for MiniMax-H3, the newest video era 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 major milestone for AI-driven video and audio era.
MiniMax-H3, launched on July 31, 2026, represents a generational leap in video synthesis. Not like its predecessor, the Hailuo collection, 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 probably the most versatile open-weight video fashions out there.
AMD’s assist 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 clean processing of the joint video-audio latents that type the core of H3’s output. That is vital for sustaining alignment between movement and sound, a trademark of H3’s capabilities.
The mannequin structure itself is very modular. MiniMax-H3 splits duties throughout two partitions: FL2VA for text-to-video and frame-conditioned enter, and Ref2VA for reference-driven era. Every partition consists of elements 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 completely different {hardware} setups.
Past technical developments, MiniMax-H3’s open-weight launch beneath the MiniMax Group License is a serious draw for builders. It lowers limitations to entry for these experimenting with video era, positioning H3 as a dominant participant within the open AI panorama. The mannequin is already out there 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 indicators its competitiveness within the AI {hardware} market, the place NVIDIA’s dominance has lengthy been unchallenged. The collaboration additionally underscores the growing demand for high-performance GPUs able to dealing with subtle diffusion-based fashions.
Trying forward, AMD’s assist for MiniMax-H3 units the stage for additional optimizations, together with batch processing capabilities and improved kernel efficiency. As AI-driven video era beneficial properties traction, each AMD and MiniMax seem well-positioned to capitalize on the rising curiosity in multimodal content material creation.
Picture supply: Shutterstock
