MiniMax H3 Max on fal: AI Video Faster Than You Can Watch It
fal's MiniMax H3 Max renders a 5-second clip with audio in under 3 seconds, and within days builders turned that into infinite AI TV channels, a 24/7 AI news stream, and prompt-driven games. I look at how it works, what people built with it over one weekend, what it costs, and whether it'll be open-sourced.
I’ve been keeping an eye on real-time AI video for a while, and my interest traces back to two quotes: LTX’s CEO, Zeev Farbman, saying at the LTX Video launch a couple of years back that video generation faster than you can prompt would be possible soon, and Jensen Huang’s now-famous line that “in the future, every pixel will be generated.” That future arrived faster than I expected, mostly because of open source: MiniMax open-sourced its H3 video model, which by my read has become the consensus favorite among open-source video builders, and fal just built a faster, hosted layer on top of it called H3 Max. The full breakdown’s in the video.
This isn’t the first time real-time video’s come up here, either — Decart’s Lucy 2.5 tackled it from the restyling side back in July. H3 Max comes at it from the generation side instead.
Who’s behind it, and what H3 Max does
fal isn’t a familiar name outside infrastructure circles, but chances are you’ve used it without realizing — it’s the inference layer a lot of gen-AI platforms run on. I’ve covered fal before and was recently on their podcast with Matt Workman. Lately they’ve started shipping their own front-end products too, and H3 Max is the first big one: built on open-source MiniMax H3, it renders a 5-second, 768p clip with audio in under 3 seconds — faster than real time.
Is faster automatically better? I don’t think so — it comes back to the old fast-cheap-good triangle, pick two. MattVidPro ran an A/B test — H3 Max against locally-generated H3 with a Turbo LoRA — and it’s close: the local version edges it out slightly, but both clips get the idea across fine (the test subject was a crab with a jetpack).
Where H3 Max wins outright is price: it’s free to try right now on fal, with a handful of free generations a day at 5 or 10 seconds, text-to-video or image-to-video. My own test — a guy in a blue suit walking down a city street — held up well. One caveat: you’re technically queued, so total wait time runs a bit longer than the 3-second generation itself; mine came in under 30 seconds.
fal Live, infinite TikTok, and Infinite Slop
The half of this that actually got me excited is what happens once you bolt interactivity onto faster-than-real-time generation. fal Live is a set of six streaming channels — Chaos Stream and Hot Mic among them — generating continuously off prompts coming in through live chat. As of this recording, fal Live itself was on standby while the team tinkers with it, but there’s a capture floating around from Yegor Sak’s session (recorded, it sounds like, on an external mic, so the audio’s thin).
Within days, people outside fal started building on the same idea: one project strings H3 Max into an infinite, TikTok-style vertical feed — random, but no more random than scrolling any other vertical feed — and another leans into the Rick and Morty “interdimensional cable” bit, which fits the randomness well.
The biggest is Infinite Slop AI, built by LevelsIO — reportedly on his phone, in a sauna — and live as I write this, with over a thousand concurrent viewers prompting and reacting together. Worth flagging the distinction from fal’s own tooling: fal Live runs on the H3 Max Director, an autoregressive variant with up to two minutes of contextual memory, built to hold consistent characters and locations across prompts. Infinite Slop has no director layer at all, and yet something like a narrative still emerges anyway: hive-mind, surreal, oddly mesmerizing to watch for a few minutes.
Channel Zero, and an interactive (very recognizable) sitcom
Yegor Sak also built Channel Zero: a 24/7 AI news channel pulling from RSS feeds, Reddit, and — genuinely — an antenna mounted to his roof, turned into prompts and then into video. He’s posted roughly two hours and forty-five minutes of it to YouTube so far. Weird to watch, but hypnotic in the same way the streams above are.
The demo I keep coming back to is an interactive sitcom bit that’s unmistakably Seinfeld — viewers vote between generations on where the scene goes next. You can see the seams: each clip is still bound by that roughly 5-second window, and the stitching between them looks like it might be a first-frame/last-frame trick. Doesn’t make it less impressive, but it’s early.
It’s also, obviously, an IP question the industry will have to work out at some point — not something this one video resolves. I did laugh at the demo poking fun at itself for constantly getting recycled as an AI benchmark, temporal-consistency jokes included. One idea from the video that stuck with me more than most: a live D&D campaign where chat votes on the party’s decisions at each turn.
A playable game, and what this actually costs
Past the passive streams, there’s a real-time video Pictionary game built on the same tech — the kind of thing that couldn’t exist without generation this fast. Blendy went further and open-sourced a branching, choose-your-own-adventure game, which reminded me of Space Ace or Dragon’s Lair (if you’re old enough to remember those, you’re probably complaining about your back a lot lately too — no judgment, same here). It’s open source, though it calls the fal API plus two small LLM calls per story beat, so budget for that if you fork it.
On pricing: as of this recording, the API looks to run about 8 cents a second at 768p — roughly $4.80 a minute — though there’s a launch sale on right now at 4 cents a second for the first 14 days. Running something like a 24/7 Infinite-Slop-style channel at that rate would cost close to $4,000 a day by my rough math — not worth it right now.
On open-sourcing: nothing official, but there’s chatter H3 Max itself might get open-sourced eventually. I wouldn’t expect the Director layer to follow — my guess is fal keeps that piece proprietary.
A slop cannon today, a new medium tomorrow
So where does this leave things? Real-time streaming AI video is genuinely here. Is the quality at Seedance levels? No, it is not. Is it, to borrow LevelsIO’s own name for his project, a bit of an AI slop cannon? Currently, yes. But quality in this category moves fast — real-time AI video looked considerably rougher as recently as nine months ago.
That’s the tension: this opens the door to things like infinite episodes of The Office — a lot of empty calories. But the piece I’m actually excited about is the interactive layer: prompted, voted-on, branching video, generated as fast as a room can type. Serious narrative work will always need intentional craft, generated or traditionally edited — but this instant, interactive form is genuinely different, and it opens narrative doors that didn’t exist before. I’ll keep watching this space, and now I’m curious what you’d build with it.
Where I land: it’s here, it’s free to try, and the quality gap won’t hold — nine months ago real-time video looked far rougher than this. A 24/7 channel isn’t worth the bill yet, and I’d expect the Director layer to stay proprietary. The interactive layer is the part I’m actually excited about, and it’s the part I’d point you at.