FLUX 3 Max is a new video generation model post-trained by fal together with Black Forest Labs. On October 19, 2026, fal announced the model in a PR Newswire release: it scored an Elo of 1,139 in human preference evaluations on the Artificial Analysis Video Arena, placing it in the top tier for image-to-video quality, and it generates a 10-second video in just 10.95 seconds. The model is available from today on fal's inference platform through a unified API covering text-to-video, image-to-video and keyframes-to-video, at 75% off for the first three weeks.
Not a little faster: a rival at the same score needs over four minutes
fal's latency comparison is blunt. At the same reported 1,139 Elo quality level, Seedance 2.5 takes 277.80 seconds to generate a 10-second video, while FLUX 3 Max takes 10.95 seconds — roughly 22 times faster. Against the original FLUX 3, its image-to-video Elo rises from 1,089 to 1,139 while generation latency falls from 126.44 seconds to 10.95 seconds, an 11.5x inference speedup. The evaluation fine print matters too: the 1,139 score comes from a preliminary evaluation across 4,868 votes, and its confidence interval overlaps with the highest-ranked models — fal's own wording is top tier, not sole number one.
Beyond the base model, post-training is the second battleground
FLUX 3 Max builds on Black Forest Labs' FLUX 3 foundation model, with fal Research adding specialized post-training, reinforcement learning and inference optimization aimed at visual fidelity, motion coherence and prompt adherence at the same time. Robin Rombach, co-founder and CEO of Black Forest Labs, said FLUX 3 was built as a foundation for others to build on; fal CEO Burkay Gur pointed at systems co-design, arguing that some of the most important advances in generative AI will come from optimizing models and the systems that run them together. fal says this is only the first iteration of FLUX 3 Max, with further post-training already under way.
What this changes for video teams
When a 10-second clip drops from a wait of more than four minutes to roughly eleven seconds, the workflow changes with it: storyboards can be tried in more variants, failed takes cost seconds, and near-real-time preview workflows become plausible. Teams evaluating it should check three things: re-test quality on their own material instead of trusting arena scores; the 75% discount lasts three weeks, so long-term costs should be estimated at full price; and image-to-video is its verified strength, so pure text-to-video performance needs separate testing. The Mercury 3 diffusion model covered on this site is running the same race on the text side: at equal quality, whoever cuts latency first wins.
