VDN-MiniMax-H3 renders 14s of video in 11s — but not in the US
Hugging Face's top trending model ships full training code and 8-GPU inference — but its license bars the US, EU, UK, and Korea. Plus the hosted route.
Copy markdown
14.4s of 768p, rendered in 11.23s
On eight B200 GPUs at 8 denoising steps, VDN-MiniMax-H3 renders a 14.4-second 768p clip in 11.23 seconds — faster than the clip itself plays. Eight H200s do it in 18.3s; a single B200 in FP8 takes 51s. It handles text-to-video, image-to-video, and first-to-last-frame interpolation, each with synced audio.
How it's fast: linear attention bolted on
Rather than retrain, OpenVDN adds a frame-wise linear-attention branch alongside MiniMax H3's softmax backbone and ships it as plug-and-play LoRA adapters — so you keep the base model's visual quality while cutting the compute. The optimized inference stack and full training code ship with the ~82GB weights.
The catch: the license locks out its own audience
The MiniMax H3 Community License bars use in the US, EU, UK, and South Korea — most of the people who would want it. And those headline speeds assume eight datacenter-class GPUs. Read the terms before wiring it into anything you ship.
What you can actually run today
The base MiniMax H3 is live as a serverless API on fal (a day-0 partner): 5–15s clips up to 2K at 24fps with native audio — score, dialogue, and foley timed to the cut. Its 'H3 Max' tier renders a 5-second 768p clip in under 3 seconds, pay-per-use. That's the door if the weights' license shuts you out.