What happened. On 27 August 2026 fal published H3 Max, a version of MiniMax's open-weights H3 video model that fal Research post-trained for prompt adherence and visual quality and that fal's inference team optimised for speed. fal states that H3 Max generates a five-second clip in roughly three seconds — around thirty-five times the throughput it measured for the official H3 endpoint — and that it leads its human-preference evaluations across overall quality, prompt understanding and aesthetics. The model is available in fal's playground and through its API; a faster Turbo variant is offered alongside it.
Why it matters. Faster-than-real-time generation is the enabling condition for continuous channels: as long as the next clip finishes before the current one ends, a channel never has to pause. fal's own framing puts it in exactly those terms, and within days of the release independent developers were running always-on channels on the endpoint. Whether the benchmark claims hold up is a separate question we do not evaluate; what we can observe is that public channels built on it are on air.
Which platforms it affects. Infinite Slop credits fal's H3 Max on its page. fal.live labels its channels with the same model in its page metadata. Both platform pages link back to the H3 Max model brief, which records only the facts fal has stated publicly and marks weights availability as not disclosed.