FLUX 3 Max's 22x speed claim shrinks to 5.5x on a second chart

fal Research and Black Forest Labs launched FLUX 3 Max on October 19, 2026, a post-trained version of Black Forest Labs' FLUX 3 video model, and the speed figures fal published for it do not agree with each other.

FLUX 3 Max ships under fal's name. fal Research took BFL's FLUX 3 foundation model, continued training it on new data and objectives, a process both companies call post-training, and then optimized how it runs on fal's inference platform. It generates video from a text prompt, a starting image or a set of keyframes.

Two of fal's own pages give different times for the same 10-second clip. The product page states a 12.45-second average and calls FLUX 3 Max about 22x faster than ByteDance's Seedance 2.5 at 277.80 seconds. The image-to-video endpoint page says approximately 11 seconds and "more than 25x faster generation" against the same 277.8-second figure for Seedance 2.5. fal's syndicated launch release, quoted by Unite.AI, uses a third number, 10.95 seconds, and calls that over 22x faster than Seedance 2.5.

TokenPost reported on October 20 that fal later published a second chart which kept FLUX 3 Max at 12.45 seconds but put Seedance 2.5 at 69.09 seconds, which works out to about 5.5x. TokenPost said both charts cover 10-second, 16:9 clips at 720p or 768p, and that fal has not explained why the Seedance 2.5 figure changed. The same report says MiniMax H3 moved from 83.26 seconds to 122 seconds and the original FLUX 3 fell from 126.44 seconds to 46.29.

The quality half of the claim sits on Artificial Analysis's blind human-preference boards, and on those the live pages back fal up. Reading AA-Video-I2V v1.1 on October 20, 2026 at 11:07 UTC with the "With Audio" toggle on, FLUX 3 Max (720p) appears second at 1,138 Elo across 5,054 samples, behind MiniMax H3 (768p) at 1,140 and level with Dreamina Seedance 2.5 at 1,138. The top three overlap in their 95% intervals, so the point estimates do not separate them. On AA-Video-T2V v2.1, read the same day with audio selected, FLUX 3 Max is fourth at 1,159, tied with third-place Dreamina Seedance 2.5. Base FLUX 3 is eighth on the image-to-video board at 1,086 and seventh on text-to-video at 1,145, so Max is ahead of it on both.

FLUX 3 Max follows H3 Max, fal Research's earlier post-trained video model, which fal says generated video in 3.49 seconds in its own benchmark. fal calls this launch the first iteration of FLUX 3 Max and says further post-training is underway. Both post-trained models are served on the same API, so a shot can move between H3 Max and FLUX 3 Max by changing a model ID.

The launch rate is $0.0425 per second of 720p output, 75% off the $0.17 list price, and it runs through November 7, 2026, with 1080p at $0.0725 per second during the promotion and $0.29 after. fal's pricing table, read the same day, lists FLUX 3 Max at $0.085 per second, a figure the product pages do not reconcile.

A post-training pass over someone else's foundation model is cheap to repeat. The figure that will settle this is the wait a buyer measures on their own prompts. If the second chart's 5.5x holds, the gap to Seedance 2.5 is roughly a quarter of the one that led the launch release.

FLUX 3 Max is live now through fal's API, with text-to-video and image-to-video endpoints at fal.ai/flux-3-max.