NVIDIA Releases Sol-H3 Inference Code, Running MiniMax H3 Video Faster Than Playback on Eight B300 GPUs
- NVIDIA's Sol-H3 generated a five-second MiniMax H3 video with synchronized stereo audio in under two seconds on eight B300 GPUs, faster than the clip's playback time.
- The FastH3 adapter cuts H3's generation path from 49 transformer forwards to four, using approximate sparse attention and INT8/FP8 data transport.
- NVIDIA measured text encoding, denoising, and video and audio decoding, but excluded model loading, compilation warmup, and final MP4 encoding from its timing.
- On a single B300, the same five-second output took 13.745 seconds, so the reported faster-than-playback result does not apply to a one-GPU setup.
- Sol-H3 code is under Apache 2.0, but H3 weights use separate terms and MiniMax temporarily restricts open-weight use in the US, EU, UK, and South Korea without authorization.