fal, the inference platform that hosts image, audio and video models for developers, launched FLUX 3 Max on Monday, October 19, 2026. The model is not a new foundation model. It starts from FLUX 3, the video model made by German lab Black Forest Labs, and adds further training and inference work by fal's own research team. Both companies promoted the release, and Black Forest Labs chief executive Robin Rombach said in fal's announcement that FLUX 3 was designed as a base for others to customize.
The pitch is speed without a visible drop in quality. In fal's own benchmark, a 10-second, 16:9 clip at 720p takes about 12.45 seconds on its product page and 10.95 seconds in its press release, against 277.80 seconds for ByteDance's Seedance 2.5 and 126.44 seconds for the original FLUX 3. fal describes that as roughly 22 times faster than Seedance 2.5 and more than ten times faster than the base model. Those timings are vendor measurements, and the API documentation notes that settings such as a higher prompt-reasoning level make generation slower.
Quality is where outside data exists. Artificial Analysis, which runs blind human-preference arenas for generative models, said on launch day that FLUX 3 Max ranks second on its AA-Video-I2V v1.1 image-to-video leaderboard, statistically tied for first, and fourth on its AA-Video-T2V v2.1 text-to-video board, ahead of base FLUX 3 on both. fal's materials cite a preliminary score of 1,139 Elo, up 50 points from FLUX 3. When VentureBeat checked the board the same day, it showed FLUX 3 Max at 1,138, matching Seedance 2.5's score and sitting two points under MiniMax H3, with overlapping confidence intervals that make the top spots hard to separate.
The model is offered only as a hosted service. fal exposes text-to-video, image-to-video and keyframe endpoints; the image-to-video schema accepts a start image and an optional end image, clip lengths from 5 to 20 seconds, 720p or 1080p output, several aspect ratios and audio generation that is on by default. VentureBeat reports that the keyframe mode accepts up to 10 images. The launch materials do not mention downloadable weights.
Pricing is per second of output. fal lists $0.17 per second at 720p but is charging $0.0425 until November 7, a 75% launch discount that puts a 10-second clip at about 43 cents; VentureBeat puts the regular 1080p rate at $0.29 per second. fal also points developers to H3 Max, an earlier model its researchers post-trained, which runs on the same API, so teams can swap between the two by changing a model ID.
For Black Forest Labs the release shows a partner building a commercial product on its video foundation model. For fal, which mostly serves other companies' models, it is a sign that the platform wants to compete on model quality and latency of its own making rather than on hosting alone. Faster generation matters most in interactive work, where designers or product teams iterate on many takes of the same shot.
Several caveats apply. Speed figures come from fal and differ between its own documents. The arena ranking is preliminary, rests on human preference votes for clips generated from images, and does not prove better performance on every production need. No independent latency tests were found, and how the model behaves at regular prices and under heavy load remains to be seen.