An AI model is now generating video faster than you can watch it. Fal engineer Rehan Sheikh set up a broadcast using H3 Max, the company’s optimized version of the MiniMax H3 video generation model, that cranks out 15 seconds of video content in roughly 9 seconds. The result is a continuous, automated stream where the AI is literally producing footage quicker than real-time playback.
What H3 Max actually does
H3 Max launched on August 27, 2026, as a closed, post-trained iteration of MiniMax H3, the open-weight multimodal model that debuted in July 2026.
The model renders a 5-second clip at 768p or 720p resolution in under 3 seconds. Some tests have clocked it at around 2.5 seconds. For longer clips, the 15-second generation window lands somewhere between 9 and 16 seconds depending on resolution and settings.
According to independent benchmarks, H3 Max delivers up to 35 times the throughput of its predecessor. It also ranks first in human preference tests across quality, aesthetics, and prompt understanding for image-to-video with audio performance.
On the pricing front, fal is running a promotional rate of roughly 12.5 cents per 5-second clip at 480p, with the cost doubling for higher resolutions.
The livestream experiment and its platform journey
Sheikh’s experiment was a proof of concept for an ongoing, automated video stream generated entirely by AI in real time. No pre-rendered assets, no human operator choosing clips.
The stream launched around August 29, 2026. Sheikh initially broadcast on Twitch, which quickly became a problem. The stream migrated to Kick next, where it ran into similar friction. Both platforms flagged the content through their moderation systems. Rumble became the stream’s landing spot, with its lighter-touch approach to content moderation making it a more hospitable environment for the experiment.
Why faster-than-real-time matters
When generation speed exceeds playback speed, use cases that were previously impractical become viable: automated content channels that never go offline, real-time personalized video feeds, or advertising pipelines that can produce bespoke creative assets on the fly without rendering queues.
The throughput gap—35x over the predecessor—suggests that architectural improvements rather than mere hardware scaling are driving the gains.
The Twitch and Kick rejections of Sheikh’s stream are early signals of a friction point that will only intensify as these tools proliferate. Rumble’s willingness to host the stream positions it as a potential home for experimental AI content.