FastH3-Omni Cuts MiniMax H3 Video Times

Turning a single reference image into a five-second clip with sound took MiniMax’s H3 model 35.44 seconds on eight Nvidia GB200 chips. A distilled version released on Tuesday did the same job in 5.67 seconds, according to the team that built it.

That team is Nuva Lab, working with FastVideo at UC San Diego’s Hao AI Lab, and the model is FastH3-Omni. It keeps H3’s ability to build a 5- to 15-second video at 24 frames per second, with synchronised audio, from a mix of reference material, but it gets there in a fraction of the computation. Across 27 test workloads, the group reports end-to-end latency 5.6 to 11.9 times lower than the original.

Eight steps instead of 50

The saving comes from distillation, training a model to reach in a few steps what its teacher reaches in many. According to the model card, FastH3-Omni uses “eight transformer evaluations”, against 50 for the base model, and accepts up to nine images, three videos and three audio clips, with no more than 12 references in total. Output is 32-kHz stereo audio at 24 frames per second.

The lab’s technical write-up shows where the range comes from. The smallest gain, 5.63 times, came with nine images and a five-second clip; the largest, 11.93 times, with one short video clip and a 15-second output. Training took about 7,800 GPU-hours on 64 Nvidia B300 chips, supplied by Reactor.

The numbers are the team’s own

These are the builders’ measurements on their own benchmark, and the write-up is careful about what they mean. The comparison “compares complete model/runtime recipes”, not identical code paths, and the base model ran under SGLang’s “lossless” setting rather than its stricter “exact” one. The authors call the gallery of side-by-side clips “qualitative comparisons, not evidence of output parity”. Until someone outside the project runs the same tests, the quality cost of the shortcut is probably the main open question.

The licence is the other catch. The weights are downloadable, but “Use of these weights is governed by the MiniMax-H3 Community License Agreement”, not a standard open-source licence, and the runtime still loads H3’s text encoder and VAEs from MiniMax. Nuva also sells a paid production version.

For people without a GB200 rack, the card lists 4-bit builds for Nvidia’s 128 GB DGX Spark and the 32 GB RTX 5090, with “Both profiles default to 480p”, plus 8-bit weights for the RTX 4090; Nuva’s page adds Apple Silicon support. Reference-guided video models with sound are arriving fast, a day after fal and Black Forest Labs launched FLUX 3 Max, which fal says makes a 10-second clip in under 11 seconds.

How fast FastH3-Omni runs on a single desktop card is the figure most users will want, and the lab has not yet published it: its post on local use is planned for 21 October.