AMD announced Day 0 support for MiniMax-H3, a video generation model from Chinese AI firm MiniMax, on its Instinct MI355X and MI300X GPUs. MiniMax-H3, released on July 31, 2026, unifies text, image, video, and audio inputs into a single framework, unlike its predecessor, the Hailuo series. The model generates up to 15-second clips at 2K resolution (1440p) with synchronized stereo audio, surpassing the 1080p, 10-second output of earlier models. It also introduces instruction-based video editing and reference-driven animation. AMD’s support relies on its ROCm platform and SGLang framework, employing 8-way sequence parallelism and leveraging AMD’s AI Tensor Engine (AITER) kernels. The model architecture splits tasks across two partitions: FL2VA for text-to-video and frame-conditioned input, and Ref2VA for reference-driven generation. MiniMax-H3’s open-weight release under the MiniMax Community License lowers barriers to entry for developers. The model is available on platforms like Atlas for $0.14 per second at 2K resolution. This launch highlights the growing synergy between AI workloads and AMD’s Instinct GPU line, signaling competitiveness in the AI hardware market. Future optimizations include batch processing capabilities and improved kernel performance.