NVIDIA and SANA Release Sol-H3 Stack That Generates 5-Second MiniMax-H3 Video in 1.653 Seconds
- Sol-H3 generated a 5-second, 1344×768 MiniMax-H3 video with stereo audio in 1.653 seconds on eight NVIDIA B300 Blackwell GPUs, below the clip's playback time.
- Compared with Base H3's 18.250 seconds and 50 DiT forward steps, Sol-H3 uses four DiT steps and reached an 11.04x speedup on the same eight-B300 setup.
- On eight B300 GPUs, Sol-H3 generated 10-second and 15-second videos in 3.732 and 6.612 seconds, with reported speedups of 13.57x and 15.05x.
- The project combines NVIDIA's Sol Engine, Sol-Attn sparse attention, and MiniMax's four-step distillation; its code is released under Apache 2.0 and Reactor provided an API at launch.
- Benchmarks used 1344×768 output at 24 FPS with stereo audio, reported the median of three runs, and excluded model loading and compilation warm-up time.