Vision Beating AI News Flash: AI inference platform fal has released H3 Max. It is a video model trained by fal based on the open-source MiniMax H3, with a focus on enhancing keyword understanding and visual effects, and has been specifically optimized for their in-house inference system.

In the Artificial Analysis Sound-inclusive Video Ranking, the H3 Max image-based video ranks first with a 1205 Elo, surpassing ByteDance's Seedance 2.0 and the original H3; the text-based video ranks third with a 1235 Elo, only 4 points behind the first-placed Wan 3.0.

The speed is also impressive. fal stated that generating a 5-second 768p video takes less than 3 seconds, with a throughput approximately 35 times that of the original H3 interface. Currently, it is launched at a 50% discount, costing only $0.04 per second for 768p, and generating a 1-minute video for just $2.4; this price will revert to $0.08 per second after September 1.

However, the capabilities of H3 Max are more limited. It currently only supports up to 768p resolution, while the original H3 can generate 2K videos; the initial release only includes text-based and image-based videos, with the Reference-to-Video feature not yet available.