[fal retrains MiniMax H3: Image-to-video tops the charts, 5-second video generated in 3 seconds] Beating AI Newsflash: AI inference platform fal has released H3 Max. It is a video model trained by fal based on the open-source MiniMax H3, with enhanced prompt understanding and visual effects, and specifically optimized for its proprietary inference system to accelerate performance. On Artificial Analysis's sound-enabled video leaderboard, H3 Max's image-to-video scored 1205 Elo, ranking first, surpassing ByteDance's Seedance 2.0 and the original H3; text-to-video scored 1235 Elo, ranking third, just 4 points behind the first-place Wan 3.0. The speed is also remarkable. According to fal, generating a 5-second 768p video takes less than 3 seconds, with throughput approximately 35 times that of the original H3 interface. Currently, the launch price is 50% off, costing only $0.04 per second for 768p video, meaning generating a 1-minute video costs just $2.4; after September 1, the price will revert to $0.08 per second. However, H3 Max has a narrower capability range. It currently supports a maximum resolution of 768p, whereas the original H3 can generate 2K videos. The launch version only includes text-to-video and image-to-video functionalities, with Reference-to-Video (reference-based image-to-video) not yet available. [Original link]