MiniMax H3 video model goes open source, renders clips faster than they play
OpenVDN has released VDN-H3, a fully open-source hybrid-attention video model built on MiniMax H3. The headline number: an 8-step, 14.4-second clip renders in 11.23 seconds on 8 B200 GPUs, faster than real time. That speed comes from a frame-wise linear attention branch grafted onto the softmax backbone via two mergeable LoRA adapters. Weights, inference stack, and training code are all public, with a single-GPU script for first renders.