VDN-H3 Makes Fourteen Seconds of Video in Nine. Then the Licence Tells You Where You Can't Run It.
A video model that renders faster than it plays, the real numbers on one card instead of eight, and the cheapest lawful way to try it this week.
A research group published a video model yesterday that finishes a fourteen-second clip, with synchronised sound, in about nine seconds. Not fourteen minutes. Nine seconds, for fourteen seconds of footage. Then you scroll to the bottom of the model page and find the licence, which says the rights it grants apply worldwide excluding the European Union, the United Kingdom, the Republic of Korea, and the United States of America.
So the fastest openly published video model of the week is one that most of the people reading about it are not licensed to run. That gap is worth sitting with, because both halves of it are real and neither cancels the other out.
The model is called VDN-H3, from a team led by Haocheng Xi at UC Berkeley with colleagues at Impossible Inc. and UT Austin. The architecture write-up and the paper both went public on September 17. The weights had been sitting on Hugging Face since September 2, which is why the release felt like it arrived sideways.
Why crossing playback speed changes the work, not just the wait
Here is the thing that matters, and it has nothing to do with benchmarks.
When a clip takes fourteen minutes, you write the most complete, over-specified prompt you can, and you hope. One attempt per coffee. You cannot afford to be curious, so you front-load every decision into text before you have seen anything.
When a clip takes nine seconds, you stop writing careful prompts. You make the obviously wrong version on purpose, look at it, change one thing, and go again. That is the loop every working editor, animator and photographer already recognises, because it is how you work with a camera, a timeline or a brush. Faster generation is not a convenience feature. It moves this whole category from commissioning to sketching, and sketching is where the good stuff comes from.
fal crossed the same line this week from the commercial side and said it more plainly than anyone: once the model runs faster than the footage plays, you get video you iterate on rather than wait for. Their platform write-up on H3 Max, out a few hours after the VDN paper, puts a five-second clip at under three seconds.
How it gets fast, without the jargon
VDN-H3 is not a new model. It is MiniMax H3, the open video-and-audio model a lot of people already run, with the part that eats the time replaced.
The team measured that more than 85% of H3's runtime goes into one operation: letting every moment of the video look at every other moment. That is how the model keeps a character's jacket the same colour in frame 300 as in frame 4. It is also brutally expensive. They note that the cost of this kind of looking grows with the square of the sequence length, which is the property their whole design works around, so my own read is that doubling the footage roughly quadruples that particular bill. They do not put it in those words, so take the arithmetic as mine rather than theirs.
Their fix is to split that looking into two jobs with different budgets.
Nearby frames get the expensive treatment. Frames close together in time still look at each other exactly and fully, which is what preserves fine detail and stops the image jittering from one moment to the next. Short-range consistency is where sloppiness shows up immediately, so they did not cut corners there.
Distant frames get a running summary. Instead of frame 300 examining frame 4 directly, the model keeps a rolling précis of everything that came before, and a second one of everything that comes after, and frame 300 reads both. Think of it as the difference between rewatching the whole film to remember a character and reading your own notes about them. The notes are cheaper and mostly good enough.
Everything also gets to see the two bookends. On top of the above, every frame is allowed to look directly at the first and last frames of the clip, and those two frames look at everything. The team measured this as adding only 3.57% more work while noticeably improving how well the clip holds together end to end. It is a small, clever addition: anchor the whole thing to its own opening and closing image, and the drift in the middle gets much less bad.
Two learned volume controls balance the exact path against the summary path so one does not drown out the other. That is the whole idea. Keep precision where the eye catches errors, approximate where it does not, and anchor the ends.
The result is a bolt-on rather than a fork. The release adds one new component plus two small add-on files that merge into the original model at render time without touching its weights. If you already have H3, you are adding roughly 5 GB to an existing 72 GB, not downloading a second model.
The numbers that actually change a decision
Here is the published table, trimmed to the rows that matter. All figures are for a 14.4-second clip at 768p, and all of them are the authors' own measurements.
On one B200, the top-end card:
- Stock MiniMax H3, the slow way: 13.95 minutes
- VDN-H3, the fast way: 51 seconds
On one H200, the previous generation:
- VDN-H3, the fast way: 90.5 seconds
On eight B200s:
- VDN-H3 on the team's own stack: 11.23 seconds
- VDN-H3 on the serving stack they recommend: 6.9 seconds of rendering, about 9 seconds to a finished file
Two caveats on those, both from the team. Every figure excludes loading the model, warming it up, and the final step that turns numbers back into a watchable video file. They say that in a live setup they would run those parts on separate machines. And the paper and the project page do not quite agree with each other: the paper's summary says 6.70 seconds for a 14.3-second clip, the project page says 6.9 for 14.4. Same work, different roundings. If you repeat a number, repeat one of theirs and say whose it is.
The honest read: the headline needs eight of the most expensive cards made. On the single card a person might rent, you are at 51 seconds. That is still a completely different working day from fourteen minutes, and it is not a stream.
Put this into practice
1. Try it free, in your browser, in about two minutes. Open akhaliq/vdn-minimax-h3-demo, a hosted copy running the fast 8-step version with sound. There is also OpenVDN's own demo. These are community-hosted on shared hardware, so you will wait in a queue and you will not see nine seconds. You will see the output quality, which is the thing worth checking first.
2. Write the prompt the way the model wants it. This is the step people skip and then blame the model. The VDN team say explicitly that you should rewrite your prompt through MiniMax's own prompt tooling before generating, and that it greatly improves quality. The prompt-writing skills are published. Five minutes there is worth more than five more attempts.
3. If you want the fast path today and you are in an excluded country, pay for it. fal serves a post-trained cousin of the same model at minimax/h3-max/text-to-video. Current posted price is $0.04 per second of 768p video, described on the page as a promotional rate at 50% off that ends September 30, after which it is $0.08 per second. So a 14.4-second clip at 768p costs about 58 cents now and about $1.15 in October. Their own example response on that page reports 2.53 seconds of render time. You are renting the speed rather than owning it, and for most people that arithmetic is not close.
4. Only then consider running it yourself. hf download OpenVDN/vdn-minimax-h3 --local-dir ckpts pulls about 82 GB. The single-card first render is one script, bash scripts/inference/8nfe_tuned_fp8.sh. The first run compiles for several minutes before anything happens, which looks exactly like a hang; later runs reuse that work. Before you start the download, read the licence.
Where this breaks
The licence is the real blocker, not the hardware. VDN-H3 inherits the MiniMax H3 Community License Agreement. Its applicable territory is worldwide excluding the EU, the UK, South Korea and the USA. The card states that use outside the territory is not authorised, and invites people in excluded regions to contact MiniMax about a licence. I am not a lawyer and this is not legal advice, but the plain reading is that if you are in any of those places, the free weights are not a free option for you. The architecture is public, the code is public, the training recipe is public, and for a large share of the audience the lawful route is still someone else's API.
The speed is a fleet property. Nine seconds is eight B200s plus a tuned serving stack. Fifty-one seconds is one B200. Ninety seconds is one H200. Pick the row that matches what you can actually rent, and set your expectations from that row.
The quality claim is "near-lossless," which is a comparison, not a guarantee. The team say output stays close to the original H3 while showing higher quality and better instruction following than MiniMax FastH3, the other fast version of this model. That is their evaluation of their own model against a rival. It might be entirely right. It is still worth generating the same prompt on the hosted demo and on stock H3 before you commit a client job to the fast version.
The published timings leave out the parts you will feel. Loading, warm-up, and writing the video file are all excluded. On a first cold run those are minutes, not seconds.
What I would watch instead of the speed number
The interesting thing is not that video got faster. Video has got faster every month for two years. It is that the fast labs have stopped building toward faster clips and started building toward a connection that stays open.
fal's newest piece serves a model over the same protocol video calls use, so your prompt changes a running stream within seconds instead of becoming a new job in a queue. Kling's team published ShotStream back in March, which generates each shot conditioned on the shots before it, at 16 frames per second on a single card, driven by prompts you type while it runs. Put those together and the unit of AI video stops being the clip.
Nobody knows what to make in that medium yet. There is no craft for it, no vocabulary, no examples to copy. If you are the sort of person who would rather be early than comfortable, that is the most interesting empty room in this field right now, and the door is a free Hugging Face Space and a text box.
If you run the demo and the output surprises you, in either direction, I want to hear which prompt did it.
Medium metadata
- Title: VDN-H3 Makes Fourteen Seconds of Video in Nine. Then the Licence Tells You Where You Can't Run It.
- Subtitle: A video model that renders faster than it plays, the real numbers on one card instead of eight, and the cheapest lawful way to try it this week.
- Tags: AI Video, Generative AI, Video Production, Open Source, Filmmaking
- Suggested hero image: a film strip where the frames near the viewer are sharp and the distant frames dissolve into handwritten notes, the first and last frames pinned with visible tacks