fal.ai's new model — a platform built for fast inference of generative models — now generates AI video faster than you can watch it.
Used to be you'd wait 2–5 minutes for a 15-second video. Now the new MiniMax H3 Max model (here's fal's official announcement) spits out those same 15 seconds in about 9. And a five-second clip at 768p resolution comes together in under 3 seconds.
Basically, fal took the open MiniMax H3, ran post-training on it, and radically optimized inference. Speed went up roughly 50x, but they didn't kill quality to get there — in the platform's own tests and on Design Arena benchmarks, the model stays near the top for image quality and prompt accuracy. The one catch: it currently caps out at 768p, not 2K.
Why it matters
Video generation stopped being a slow batch job and turned into a realtime loop: generate, watch, immediately queue up the next one.
Indie hacker Pieter Levels (here's his profile with the demo video) jumped on this immediately and built Infinite Slop — an endless interactive stream whose name is a direct nod to the AI slop boom on YouTube. Viewers type into chat, and the model generates the next chunk on the fly, trying to keep it logically connected to the last one. As co-creator Rehan notes, the stream pulled in 37,000 views on day one. Current generation limit: four 15-second clips per minute.
If you remember, something like this already existed
In December 2022, a project called Nothing, Forever launched on Twitch — an endless AI sitcom in the spirit of Seinfeld (TechCrunch and Vice both covered its launch in detail). GPT-3 wrote the scripts, Azure synthesized the voices, and low-poly Unity rendered the picture. Four characters chatted in an apartment 24/7. At its peak in February 2023, roughly 15,000 people were watching this absurdity at once.
The sitcom's creators were selling exactly the same idea back then: why shoot 7 seasons when you can run an endless broadcast? But back then nobody could generate video faster than it plays back — so we were watching jerky 3D puppets over generated text, not live clips. We've covered how fast AI video has progressed since then in a separate breakdown.