MiniMax h3 leaps into real-time video generation with vllm-omni and fasth3

MiniMax has unveiled vLLM-Omni and FastH3, two breakthrough technologies that enable its H3 video large model to deliver real-time services. The vLLM-Omni architecture significantly reduces latency through system-level optimizations, including long-sequence attention, fused DiT operators, and parallel VAE decoding. With an eight-card B300 setup, this solution cuts response latency by 30.8% compared to Diffusers. Meanwhile, FastH3 slashes DiT forward computations from 49 to 4, achieving real-time performance in generating 10.125-second MP4 videos in 8.678 to 8.710 seconds. The system also meets real-time benchmarks for 5, 10, and 15-second videos. MiniMax’s H3 service architecture also enhances scalability with distributed offloading, encoder separation, and attention acceleration. These advancements mark a significant step toward commercializing high-quality AI video generation. The technical teams have provided production deployment guidance and outlined future directions, emphasizing the rapid progress in AI video generation. The VSA variant of FastH3 further taps into hardware potential, offering additional speed improvements. These innovations position MiniMax at the forefront of real-time video generation, addressing long-standing latency challenges in the industry.