OpenVDN's VDN-MiniMax-H3 Generates 14.4s of 768p in 11.23s

OpenVDN's VDN-MiniMax-H3 checkpoint renders 14.4 seconds of 768p video in 11.23 seconds on eight B200 GPUs at eight denoising steps. The model card on Hugging Face puts it in one line: this "generates video faster than it plays."

Instead of retraining the MiniMax H3 backbone, VDN bolts on a second attention path. The card describes "one frame-wise linear attention branch that is highly efficient, and a softmax branch that maintains the backbone's visual quality and consistency." Two small LoRA adapters ride with it, and OpenVDN says they "can be merged into the backbone during inference without touching the backbone weights."

Without the eight-GPU stack the numbers move. A single B200 takes 51 seconds at eight denoising steps and 5.3 minutes at 50. Eight H200s land in between, at 18.3 seconds for the same clip length.

The team says it does not "just open-source the weights," releasing the optimized inference stack and training code alongside. The catch sits in the license. MiniMax H3 Community terms scope distribution to "worldwide excluding the European Union, the United Kingdom, the Republic of Korea, and the United States of America," cutting out most of the obvious commercial audience for a text-to-video model. The release lands into a busy stretch for open video generation; SolarWM shipped its own open weights and data three days earlier.