fal launches FLUX 3 Max for three video API workflows

fal launched FLUX 3 Max, making the post-trained video model immediately available through its API for text-to-video, image-to-video and keyframes-to-video generation. Developers can use all three live endpoints without waiting for a later access rollout.

The release builds on FLUX 3, a foundation model built by Black Forest Labs, whose related FLUX 3 image-model work provides context for the product line. fal says its research team used specialized post-training, reinforcement learning, inference optimization and serving-path engineering instead of training a new foundation model from scratch. The text-to-video endpoint offers low and high reasoning settings, with the higher setting increasing generation time. The image-to-video and keyframes-to-video endpoints were also live at the research cutoff, although the main product-page FAQ mentions only the first two launch modes.

fal claims the model takes an average of 12.45 seconds to generate a 10-second 16:9 clip at 720p or 768p. That compares with 277.80 seconds for Seedance 2.5, making FLUX 3 Max about 22 times faster by fal’s account. The company’s syndicated launch release instead reports 10.95 seconds for FLUX 3 Max against the same comparison time. The supplied materials do not reconcile the two figures or disclose a reproducible test protocol, hardware configuration, queue treatment, sample count or variance. No independent workload test was available in the opened evidence.

fal also cited a preliminary Artificial Analysis image-to-video score of 1,139 Elo across 4,868 votes. The opened AA-Video-I2V v1.1 leaderboard later showed 1,138 Elo from 5,054 samples, with a 1,123–1,153 confidence interval and a possible rank range of first to fifth. MiniMax H3 had the highest point estimate at 1,140. The higher sample total is consistent with a later snapshot of a live leaderboard, but no opened change log establishes the cause of the one-point movement. Because the benchmark measures blind pairwise preferences for image-to-video clips with audio, it does not establish leadership in text-to-video or every production workflow.