fal's H3 Max tops the image-to-video leaderboard
Post-trained on the open-weights MiniMax H3, the model ranks first on Artificial Analysis's image-to-video board. It generates a 5-second 720p clip in under 3 seconds and is 50% off for a week.
Infrastructure company fal announced H3 Max, a video generation model. It was not trained from scratch: fal Research post-trained it on top of MiniMax H3, a model released with open weights.
On the independent evaluator Artificial Analysis's image-to-video leaderboard, H3 Max ranks first with audio at 1203 Elo. fal says the model places first for overall quality, prompt understanding and aesthetics in both its own and third-party evaluations.
The company's second claim is speed: a 5-second 720p video generated in under 3 seconds. Artificial Analysis measures average latency at 3.49 seconds; other models compared on the same board range from 26 seconds to 176 seconds.
On price, the model costs $3.60 per minute — the lowest in the board's top fifteen. fal also says H3 Max is 50% off for one week; since the discount is time-limited, it is unclear whether the price advantage will hold.
MiniMax H3, the base model, was published with open weights, which is what allows third parties to post-train and host it. H3 Max is one result of that.
For details see fal's page and the Artificial Analysis leaderboard.
