NVIDIA and MiniMax released Sol-H3, which generates five seconds of video in 1.653 seconds
NVIDIA Research (corroborated by Enze Xie on X)·medium signal
On an 8x B300 Blackwell system Sol-H3 produces 5s of 1344x768 video with stereo audio in 1.653 seconds using 4 DiT forwards, against 18.25 seconds and 50 scheduler points for base H3, an 11.04x full-profile speedup that rises to 15.05x on 15-second outputs. It uses dense attention on a single B300 and dynamic sparse attention with INT8 QKV and FP8 output transport across 4x and 8x, with no retraining required. The code is Apache 2.0 and any MiniMax-H3 few-step LoRA plugs into the same engine.