Fal Launches H3 Max Video Model, Enhancing Speed and Quality for Content Creation

On August 27, Fal unveiled H3 Max, a post-trained variant of the MiniMax open-source H3 video model. This new model boasts an impressive throughput, generating a 5-second, 768p video clip with synchronized stereo audio in under 3 seconds. According to Fal, H3 Max is approximately 15 times faster than leading competitors while maintaining similar quality levels.

In evaluations, H3 Max achieved an Elo rating of 1,341 on the Design Arena Image-to-Video leaderboard and 1,201 on the Artificial Analysis Image-to-Video with Audio leaderboard, making it a top contender in both categories. The MiniMax H3 team endorsed Fal's enhancements, noting that they enabled practical video generation at a larger scale without compromising the core capabilities of the original model.

Fal's approach involved extensive post-training focused on prompt adherence and visual fidelity, complemented by a specialized inference engine optimized for video generation. Although H3 Max operates at a lower resolution of 768p compared to MiniMax’s original 2K capability, it was specifically designed for faster-than-real-time applications such as live streaming and on-the-fly video creation for social media platforms.

To promote H3 Max, Fal is offering a 50% introductory discount for the first week, accessible through its platform and API. The original MiniMax H3 model can produce 15-second clips at 2K resolution, accommodating various input types including text, images, videos, and audio.