H3 Max by fal - fal's post-trained MiniMax H3 for quality video production

H3 Max is fal's post-trained take on MiniMax H3, same open-weight base model, but retrained for better prompt adherence and visual quality, then tuned hard for speed.

The problem it's solving: Video gen usually makes you pick one. Want quality? Wait longer. Want speed? Accept worse output. H3 Max tries to break that tradeoff instead of accepting it.

How: fal Research fed it new training data focused specifically on prompt understanding and aesthetics, while their inference team rebuilt the serving stack around it (running on NVIDIA GB200 NVL72) so the speed gains don't eat into quality.

Does it work? They ran it against 12 other video models, including the original H3, Gemini Omni Flash, Wan 3.0, Seedance 2.5, Kling 3, Veo 3.1, using human preference scoring (Bayesian Elo). H3 Max came out #1 on overall quality, prompt understanding, and aesthetics, and won most head-to-head matchups it was put in. Independent benchmarks (Artificial Analysis, Design Arena) back this up too.

Key features:

• Text-to-Video and Image-to-Video endpoints

• 5-second clip in about 3 seconds, that's 35x the throughput of the official H3 endpoint

• Usable via Playground, fal Agent, or API

Good fit if you're building anything that needs fast, high-quality video generation at scale, not just a one-off demo. Try it: Text to Video · Image to Video

P.S. I hunt the latest and greatest launches in tech, SaaS and AI, follow to be notified → @rohanrecommends

3 seconds for a 5s clip is wild if the quality holds up at that speed, most "fast" video models I've tried get noticeably worse motion coherence past 2-3x realtime. gonna try the image-to-video endpoint on something with actual camera movement, that's usually where speed tricks fall apart