Photo: Tima Miroshnichenko / Pexels
The generative media platform's new H3 Max model produces a five-second video clip in under three seconds, making AI video generation faster than real-time filming.
Fal unveiled H3 Max on August 27, a post-trained variant of MiniMax’s open source H3 video model that delivers roughly 35 times the throughput of MiniMax’s official endpoint.
H3 Max generates a 5-second, 768p video clip with synchronized stereo audio in less than 3 seconds of wall time. The model is approximately 15 times faster than leading competitors at similar quality levels, according to Fal.
In human preference evaluations, H3 Max holds 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, placing it as the number one contender in both categories.
The MiniMax H3 team endorsed the results, noting that Fal’s enhancements made practical video generation possible at broader scale without stripping out any core capabilities from the base model.
Fal’s approach involved extensive post-training focused on two areas: prompt adherence and visual fidelity. From there, a specialized inference engine optimized the video generation pipeline.
AI, tech, and the markets they move—in one daily briefing.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
H3 Max operates at 768p resolution, lower than MiniMax’s original model, which can output at 2K. Fal built H3 Max specifically for faster-than-real-time applications, including live interactive streaming, real-time content generation during broadcasts, and on-the-fly video creation for social platforms.
Fal launched H3 Max with a 50% introductory discount for the first week, accessible through both its platform and API.
MiniMax’s original H3 model was designed to accommodate various input types, including text, images, videos, and audio, capable of producing 15-second clips at 2K resolution with native stereo audio output. Fal, which had initially hosted MiniMax’s models, took the open weights and conducted post-training using new datasets focused on prompt adherence and aesthetic qualities, emphasizing reinforcement learning techniques.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our