fal's post-trained MiniMax H3 Max takes the top image-to-video slot
- Source
- x.com
- Date
MiniMax H3 Max, a post-trained version of MiniMax H3 developed by fal, debuts at #1 in Image to Video and #3 in Text to Video on the Artificial Analysis Video Leaderboards with Audio, ahead of the base MiniMax H3 on both MiniMax H3 Max is built and served by fal, and post-trained from MiniMax H3. fal describes it as being tuned for stronger prompt adherence and better aesthetics, co-optimized with their custom inference stack for higher throughput. It generates 5 to 15 second clips with native audio at up to 768p. In the Artificial Analysis Video Arena, H3 Max ranks #1 in Image to Video with Audio, narrowly ahead of ByteDance's Dreamina Seedance 2.0 720p. It ranks #3 in Text to Video with Audio, on a board where the top three models sit within 6 points of each other. fal prices MiniMax H3 Max at $0.04 per second of 768p video ($2.40 per minute). The base MiniMax H3 endpoint on fal is $0.06 per second at the same resolution. fal has stated its intent to release the weights for MiniMax H3 Max. If it does, H3 Max would become the highest ranked open weights model on both boards, ahead of MiniMax H3, which leads on open weights today. Congratulations to @fal on the release! See below for comparisons between MiniMax H3 Max and other leading models in the Artificial Analysis Video Arena 🧵
Text to Video (With Audio) Prompt [1/2]: A clear glass of water sits on a table in front of a piece of paper with blue text. The camera moves side to side; the text seen through the water should be magnified and distorted, while the text beside the glass remains normal.
Text to Video (With Audio) Prompt [2/2]: A hand catching a ball, fingers spreading on impact, wrist absorbing shock, then grip tightening. All in slow motion. Audio: Ball smack.
Image to Video (With Audio) Prompt [1/2]: Hand writes "Welcome to Artificial Analysis" on the chalkboard. Chalk scratching and soft tapping.
Context
H3 Max is a further-trained version of MiniMax's H3 developed and served by fal. The developer's release account describes work on both model training and the system that runs it, making this a distinct variant of the base model.
In its August 26 results, Artificial Analysis placed H3 Max first for image-to-video and third for text-to-video on its boards with audio. These rankings reflect viewer preferences in the evaluator's arena. The top three text-to-video models were within six points, so the reported ordering alone does not establish a decisive quality gap.
Checking sign-in…
Loading comments…










