FLUX 3 Max
fal / Black Forest Labs
A video model post-trained by fal Research on Black Forest Labs' FLUX 3: a model host shipping a frontier video model of its own, built on someone else's base. It generates 5- to 20-second clips with audio from text, an image or up to ten keyframes, at 720p with optional upscaling to 1080p, and makes a 10-second clip in about 12 seconds: not yet real time, but fast enough to try many takes. Black Forest Labs, which built FLUX 3 for customization, called it "fast enough to try anything."
The headline claims needed reading closely. "Tied for #1" in image-to-video on Artificial Analysis meant 1,138 Elo against MiniMax H3's 1,140 and Seedance 2.5's 1,138, inside overlapping confidence intervals, and "22× faster than Seedance 2.5" shrank to about 5.5× once fal updated its own chart. It was available only through fal's API, with no open weights, at 75% off for the first three weeks. With HeyGen Video 1 built on MiniMax H3 the same month, October 2026 was when post-training a strong base became a way to ship a video model.