How fal Turned an Open-Weight Model Into the Fastest AI Video Tool Available
fal's H3 Max, a post-trained version of MiniMax's open H3 model, now ranks #1 on independent leaderboards and runs free.
fal released H3 Max, its own post-trained version of MiniMax's open-weight H3 video model, and it's now ranking first on independent leaderboards for quality while generating clips dramatically faster than the model it's built on. It's also free to try right now, no account required.
What fal Actually Built
MiniMax open-sourced H3 in early August, a 33-billion-parameter model that generates video with native audio from text, images, or reference clips. fal took that open base model and post-trained it further, adding substantial new training data focused specifically on prompt adherence and visual quality, then co-optimized the inference pipeline around the result.
The output: a 5-second, 768p clip with synced audio generates in under 3 seconds, which fal says is roughly 35 times the throughput of MiniMax's own official H3 endpoint, and about 15 times faster than any other model of comparable quality. fal frames this as directly challenging the usual assumption that higher video quality requires slower generation.
Where H3 Max Actually Ranks
This isn't just fal's own marketing claim. Artificial Analysis, an independent benchmark tracker, ranks H3 Max first among video models with audio, listing it under its internal designation "MiniMax H3 Turbo (768p)." Design Arena has it ahead of the base H3 model and every other model on its image-to-video leaderboard. In fal's own internal testing, H3 Max won the majority of head-to-head matchups against every model tested, including the original H3 it was built from.
The Real Tradeoff: Speed for Resolution
H3 Max caps out at 768p rather than the 2K the base H3 model can produce. That's a structural limitation, not an oversight: MiniMax's H3 pipeline uses a separate upscaling stage to reach 2K, and MiniMax only open-sourced the base 768p model, not that upscaling component. Since H3 Max was built entirely from the open weights, there's no 2K stage available to inherit. Text-to-video and image-to-video modes are live now; a reference-to-video mode for maintaining consistent characters across shots is announced as coming later.
What's Actually Free
Anyone can generate five 768p clips a day directly on fal's tool page with no account, no sign-up, and no login. Creating a free fal account adds five more generations a day, at up to 15 seconds instead of 5. This is a genuinely low-friction way to test a top-ranked video model before committing to a paid API plan.
Correcting a Detail Circulating Online
Posts spreading on X have claimed fal confirmed it will be "open sourcing H3 Max" itself. That doesn't hold up against what's actually been published. What's confirmed is that the base MiniMax H3 model was open-sourced by MiniMax weeks before H3 Max existed. H3 Max is fal's own proprietary post-training work sitting on top of that open base model, and nothing in fal's announcement indicates H3 Max's weights will be released separately.
The Signal in the Noise
This is a meaningful data point in the open-weight AI video conversation. MiniMax released the base model, and within weeks, a third party built something that outranks it on quality while running dramatically faster, using only the publicly available weights as a starting point. That's the actual promise of open weights playing out in real time, not community tooling built around a model, but a genuinely superior product built from one.
Does knowing H3 Max is a third-party enhancement rather than an official MiniMax product change how you'd think about relying on it for real work?
Specs & Pricing
- Model: MiniMax H3 Max, post-trained by fal on top of open-weight MiniMax H3
- Speed: 5-second 768p clip with audio in under 3 seconds
- Claimed performance: ~35x throughput vs. official MiniMax H3 endpoint, ~15x faster than comparable-quality models
- Resolution: 768p only (2K unavailable, upscaling stage not part of the open-weight release)
- Modes live now: text-to-video, image-to-video
- Coming later: reference-to-video
- Free tier: 5 generations/day, 5-second clips, no sign-up, directly on fal's tool page
- Signed-in tier: 5 additional generations/day, up to 15-second clips
- Independent rankings: #1 with audio on Artificial Analysis; #1 on Design Arena's image-to-video board
