fal releases H3 Max, its first video model

On 26 August 2026, fal — a company known mainly for hosting other developers’ generative-media models through its API rather than building its own — released H3 Max, described as its first in-house video model. It is not trained from scratch: H3 Max is a post-trained version of the open-weight MiniMax H3 (Hailuo 3.0) base, which MiniMax had released weeks earlier. fal said it added new data in post-training to sharpen prompt adherence and aesthetics, and built the model around the inference engine it has spent several years optimising for diffusion models.

The pitch is speed and price rather than a new capability ceiling. fal said H3 Max renders a five-second 768p clip in under three seconds — which it put at roughly 35 times the throughput of MiniMax’s official H3 endpoint — with natively synchronised audio carried over from the base model. It priced the model at $0.06 per second of 768p video, discounted by half for the first two weeks, with a free tier of five short generations a day, and claimed it was the highest-quality, fastest and cheapest option for general video generation.

Those rankings are so far fal’s own or drawn from early public leaderboards — the company cited a first-place finish on Artificial Analysis’s image-to-video-with-audio board and a top place on Design Arena — and independent, like-for-like comparisons remained thin at release. The more durable point is strategic: a media-inference platform post-training an open model into a branded, first-party product blurs the line between the marketplaces that serve models and the labs that make them.