Most “faster” model releases shave off a modest percentage. FastH3, released by the FastVideo team at Hao AI Lab (UC San Diego) with NVIDIA and Nuva Lab, claims something closer to an order of magnitude: a 14.38x speedup on 8 NVIDIA B200 GPUs, enough to generate a 15-second, 768p video-with-audio clip in about 13 seconds. It’s also, as of this week, one of the fastest-growing downloads on Hugging Face — up 113% day-over-day (24,402 to 52,063 downloads) — which is what put it on our radar in the first place.
How it gets there
FastH3 is a distilled version of MiniMax’s H3 video-and-audio model, and the speedup comes from two techniques stacked together. The first is Distribution Matching Distillation (DMD2), which collapses what would normally take 49 transformer sampling calls down to just 4. The second is Video Sparse Attention (VSA), which exploits the fact that video-to-video attention is highly redundant — FastH3 prunes it to roughly 90% sparsity — while keeping the text and audio attention pathways fully dense, since those carry less redundant information and are more sensitive to being pruned.
Combined, the two techniques scale from an 8.16x speedup on a single B200 up to 14.38x across 8 GPUs, according to Hao AI Lab’s own benchmarks. The model weighs in at 147.9GB, and running it in anything close to real time still assumes multi-GPU, data-center-class hardware — this isn’t a laptop-friendly release, and there’s no quantized version yet.
The license complication
The part of this story that’s more interesting than the speed number: MiniMax H3’s underlying Community License excludes the US, EU, UK, and South Korea from local deployment. FastH3 inherits that restriction. A hosted demo on Hugging Face remains reachable from anywhere, so the model isn’t functionally blocked for most users — but anyone in one of those regions who wants to actually run the accelerated weights themselves is out of luck. Coverage of the release from CreativeAIShow flagged Wan 2.2’s TI2V-5B (Apache 2.0) as the closest unrestricted alternative for creators who need to self-host in a covered region.
It’s a strange split: a fully open-source acceleration recipe — the code, the distillation method, the sparse attention implementation — wrapped around a base model whose license draws a geographic line around who gets to actually deploy the result locally. For an open-weight release, that’s a real caveat worth knowing before you plan a workflow around it.