fal's H3 Max AI Video Model Renders Clips Faster Than Real Time, Twitch and Kick Both Booted the Demo Stream

An AI company just built a video model that outpaces its own broadcast.

fal, an inference platform that until now mostly hosted other labs' AI models, announced H3 Max on X on Wednesday, August 26, 2026, with a full blog post and product page following the next day. The company describes it as its first in-house video model, though it's not built from scratch. It's a post-trained version of MiniMax H3, the open-weight video model MiniMax released at the end of July 2026.

The Numbers fal Is Selling

fal says H3 Max renders a 5-second clip at 720p or 768p in under 3 seconds, sometimes as fast as 2.5 seconds. The company puts that at roughly 35 times the throughput of MiniMax's official H3 endpoint. In human preference testing, fal claims H3 Max ranked first against 12 competing video models across quality, prompt understanding, and aesthetics, and cites a first-place finish on Artificial Analysis's image-to-video leaderboard and a top spot on Design Arena.

Those numbers carry an important caveat. As the tech blog It Does What Now noted, the rankings and speed multiples are fal's own self-reported benchmarks, and independent, apples-to-apples comparisons were thin at launch. AI reviewer Julian Goldie made the same point after testing the model for his own content pipeline: treat the claims as vendor numbers until outside testing catches up.

fal says the model was trained and served entirely on Nvidia's GB200 NVL72 systems, which the company says deliver up to twice the per-chip performance of the previous generation of chips it used to train and serve earlier models. That's a hardware story as much as a software one.

Pricing claims varied across coverage. fal's own blog post and It Does What Now report a rate of $0.06 per second of 768p video, cut in half for the first two weeks, with a free tier of five short generations a day. Separate coverage from Crypto Briefing and KuCoin, which ran nearly identical text, cited a promotional rate of roughly 12.5 cents per 5-second clip at 480p, with cost doubling at higher resolutions. Those figures cover different resolutions, so they aren't necessarily contradictory, but the discrepancy is real and unresolved in the available reporting.

A Stream That Outruns Its Own Playback

Rehan Sheikh, a fal engineer, built a livestream using H3 Max that generates 15 seconds of AI video in roughly 9 to 16 seconds depending on settings. Because generation can run faster than playback, the broadcast is continuous and fully automated. No pre-rendered clips, no human picking footage, just a model producing content as fast as it can be watched, or faster.

The stream went up around August 29, 2026. Sheikh first put it on Twitch. That didn't last. The stream moved to Kick, where it ran into the same kind of pushback. Both platforms' moderation systems flagged the content, according to reporting from Crypto Briefing and KuCoin. The broadcast ended up on Rumble, whose lighter-touch moderation approach made it a workable home for the experiment.

The Fair Concern on Both Sides

If a model can generate video faster than a human can watch it, content moderation systems built to review flagged clips after the fact are structurally behind. Twitch and Kick flagging an automated, human-less stream isn't unreasonable caution, it's platforms trying to keep pace with a tool that can outproduce their review process. That's a real design problem.

At the same time, nothing in the available reporting says what specifically Twitch or Kick's systems flagged, whether it was a copyright concern, a synthetic-media policy, or something else. Neither platform has issued a public statement explaining the decision in the sources reviewed here. Rumble hosting the stream isn't evidence of wrongdoing by the other two platforms, and it isn't evidence the stream was harmless either. It's simply the platform willing to let an unreviewed, AI-only broadcast run.

What Comes Next

fal says Hermes Agent's v0.20.6 release already added MiniMax H3 Max as a selectable video option, meaning the model is being wired into automated workflows beyond fal's own demo. The company's discounted pricing window and free tier run for a limited period before rates reset to standard levels. Whether Twitch, Kick, or other major platforms issue formal policies for fully automated, faster-than-real-time AI streams, rather than case-by-case flags, remains an open question none of the current reporting answers.