What it actually is
FreeVideo showed up on GitHub in mid-September and has been picking up attention this week, with a wave of creator demos claiming MiniMax H3 generation on cards as modest as an RTX 4060. The engine ships as a ComfyUI plugin with a Windows one-click launcher and a Linux command-line tool. On paper, that makes it one of the friendliest paths yet to local H3: download, point it at a ComfyUI install, click install, and generate.
The lever underneath is not mysterious. FreeVideo builds on VDN-H3, an 8-step distilled variant of the H3 model from the OpenVDN project, combined with Video DeltaNet's hybrid attention. Distillation trades sampling steps for speed, and FreeVideo layers on the standard low-memory toolkit: weight streaming, asynchronous prefetching, chunked computation, and per-GPU FP8 compute paths that fall back gracefully on older architectures.
The numbers, read carefully
The figure doing the rounds is a 10-second 768p clip in 1 minute 49 seconds on an RTX 5090, against 3 minutes 36 in stock ComfyUI and 9 minutes for "H3 dense." Nearly five times faster sounds dramatic until you notice what is being compared. A distilled 8-step model against a full-step dense run is not the same workload; of course the distilled one wins the stopwatch. What is missing from every demo we have seen is the quality side of the trade. Faster is only interesting if the output still holds up, and nobody outside the project has published a side-by-side that lets you judge.
That does not make FreeVideo uninteresting. It makes the benchmarks marketing. The engine itself is verifiable: public repo, 28 commits, a real team listed, documentation that explains the adaptive execution planner, and community LoRA support. But treat the "5x" figure the way you would treat any vendor's slide.
One more footnote
The code is Apache 2.0. The model weights are not. FreeVideo's own README points to the MiniMax H3 Community License, which carries territorial and acceptable-use restrictions — the same class of fine print we flagged when H3's license terms surfaced earlier this week. If you are building anything commercial on top of this, the license is the part to read first, not the benchmark chart.
Sources
- [1] FreeVideo repository (FlashML-org), README and license detailsRead source
- [2] OpenVDN / VDN-H3, the distilled model FreeVideo builds onRead source
- [3] Creator demos of FreeVideo's local H3 generation circulating on Instagram and Facebook, October 3–4, 2026 (benchmarks shown are self-reported by promoters, not independently verified)