fal releases H3 Max video model
fal post-trained the open MiniMax H3 weights into H3 Max for text-to-video and image-to-video. fal says it runs faster than real time and costs $0.08 per second at 768p after the launch promotion.
What it does
H3 Max is fal's post-trained version of the open-weights MiniMax H3 model for text-to-video and image-to-video. fal trained it with reinforcement learning for better prompt following and visual quality, and built a matching inference engine.
fal says it makes a 5-second video in about 3 seconds, roughly 35x the throughput of the official MiniMax H3 endpoint.
Results and limits
At launch, Design Arena ranked H3 Max #1 on its image-to-video leaderboard (Elo 1,341, just ahead of MiniMax H3 at 1,333). Artificial Analysis ranked fal's H3 model #1 on image-to-video with audio. In fal's own tests against 12 models it ranked first on quality, prompt understanding and aesthetics.
These are launch-day rankings and fal's own testing.
Worth knowing
It's available through the fal API, Playground and fal Agent. After the launch discount ended September 7, 768p costs $0.08 per second ($0.80 for a 10-second clip).
Try it yourself
Image to video: upload [still frame from your storyboard]. Prompt: "[Subject] [action], camera [move], [mood/lighting], 10 seconds." Generate several variations and keep the best take.