Fal released the H3 Max model on August 27, which is a post-trained version of the open-source MiniMax H3 video model. The new variant provides approximately 35 times the throughput of the official MiniMax endpoint and operates about 15 times faster than competing models at comparable quality levels.
The H3 Max model generates a five-second video clip at 768p resolution with synchronized stereo audio in under three seconds. While the original MiniMax H3 model supports 2K resolution and 15-second clips, the H3 Max version is specifically optimized for real-time applications like live streaming and social media content creation.
Fal achieved these performance gains through extensive post-training focused on visual fidelity and prompt adherence, alongside a specialized inference engine. The model currently holds an Elo rating of 1,341 on the Design Arena Image-to-Video leaderboard and 1,201 on the Artificial Analysis leaderboard.
The company introduced the model with a 50% discount for the first week of availability via its platform and API. The MiniMax team confirmed that these enhancements allow for large-scale video generation without compromising the core capabilities of the base model.
