fal beat Google, Kling, and Veo 3.1 at video AI. The model they used is 35x faster than the competition. Most people haven’t heard of it.
Meet H3 Max: a post-trained variant of MiniMax’s open H3 model, built and tuned by fal Research. It launched in August 2026 and immediately took the top spot on Artificial Analysis’s image-to-video leaderboard (Elo: 1,201) and Design Arena (Elo: 1,341), ranked against 12 leading models including Gemini Omni Flash, Kling 3, and Veo 3.1.
What fal actually did
H3 is MiniMax’s open-weights video model: capable, but not optimized for the quality-per-second tradeoffs that production use requires. fal took it, post-trained it with a focus on prompt adherence and visual quality, and co-designed a custom inference stack built on 4 years of diffusion infrastructure experience (fal launched in 2022).
The result: a 5-second, 768p video in under 3 seconds. That’s 35x the throughput of the official MiniMax H3 endpoint, and 15x faster than comparable-quality alternatives. The training ran on NVIDIA GB200 NVL72 systems, and the optimization work only advanced when quality evaluations stayed competitive, meaning fal didn’t trade quality for speed. They got both.
Why this matters
Most AI video tools force a choice: fast or good. Fast models spit out blurry motion artifacts that look like a dream sequence from a 2009 music video. Good models take 3-5 minutes per clip and cost accordingly. H3 Max collapses that tradeoff.
It supports Text-to-Video and Image-to-Video right now. Reference Generation (locking a character or subject across shots) is coming. Audio generates in the same pass as the video, no separate step. The human preference evaluations rank it #1 across three independent dimensions: overall quality, prompt understanding, and aesthetics.
Who should care
If you’re generating AI video for any reason (content, ads, demos, social) H3 Max is the current benchmark to test against. It’s available without an enterprise contract or waitlist. Sign up on fal, write a prompt, pick your resolution (up to 768p), choose your duration (up to 15 seconds), generate.
Pricing is $0.08 per second at 768p ($4.80/min). There’s a 50% introductory discount running right now. fal also gives you 5 free 15-second 768p generations per day to try it without spending anything.
The bigger picture
What fal did here is something worth watching beyond this one model. They took an open-weights base (MiniMax H3), ran their own post-training and inference optimization, and shipped a hosted product that outperformed everything the major labs have released in the video space.
That’s the playbook: open base + specialized tuning + fast inference = competitive product. It’s the same move that made some of the strongest models in coding and reasoning. It’s now hitting video.
The closed-model labs are not the only ones building the frontier anymore.
Try H3 Max at fal.ai/minimax-h3-max.
Chris Meredith writes about AI, technology, and what it actually means for real people. Follow along on Substack: monkeyattack.substack.com
