fal Launches H3 Max: Post-Trained Video Model Generates 5-Second Clips in Under 3 Seconds

fal has announced H3 Max, a post-trained and inference-optimized derivative of the open-weights MiniMax H3 video model, developed by fal Research in close collaboration with the MiniMax H3 team. According to fal's human preference evaluations, H3 Max ranks first in overall quality, prompt understanding, and aesthetics against leading video generation competitors — while producing a 5-second video in under 3 seconds, roughly 35× the throughput of the official MiniMax H3 endpoint and on average 15× faster than comparable-quality alternatives.

The performance gains stem from co-designing the inference engine alongside the model during post-training, rather than optimizing each in isolation. fal's team introduced substantial new training data focused on prompt adherence and visual quality, and conducted continuous head-to-head checkpoint evaluations across three independent dimensions to avoid overfitting to a single aggregate score. The result, the company says, challenges the conventional assumption that higher video quality must come at the cost of slower inference.