Fal Debuts FLUX 3 Max Video Model Post-Trained With Black Forest Labs

fal announced FLUX 3 Max on October 19, 2026, a video generation model post-trained by fal Research in collaboration with Black Forest Labs on the FLUX 3 foundation model. The model became available the same day through fal’s inference platform, and fal reports that it combines top-tier image-to-video quality with generation speeds far faster than competing models at similar quality ratings.

How Fal Research and Black Forest Labs Built the Model

The launch pairs two research teams with distinct roles: Black Forest Labs, which developed the FLUX 3 foundation model, and fal Research, whose work spans post-training, reinforcement learning, inference optimization, and systems engineering. fal Research worked directly with the FLUX 3 model, continuing its training to improve visual fidelity, motion coherence, prompt adherence, and overall generation performance.

In the announcement, Black Forest Labs co-founder and CEO Robin Rombach said FLUX 3 was built as a foundation for others to build on, and he described fal’s customization of the model for developers and enterprises on fal’s platform as the kind of collaboration FLUX was designed for. fal CEO Burkay Gur said the launch combines Black Forest Labs’ foundation model with his team’s research into quality and inference performance, adding that he believes some of the most important advances in generative AI will come from optimizing models and the systems that run them together.

Black Forest Labs’ homepage carries a banner stating that FLUX 3 Max is 22 times faster than Seedance 2.5 at the same quality, and it directs visitors to fal’s product page.

fal’s FLUX 3 Max product page describes three layers of work behind the model: post-training conducted with Black Forest Labs; inference optimization that changed how the model executes to cut generation latency and raise throughput while preserving quality; and systems engineering that tuned the serving path on fal’s own inference engine for production workloads.

FLUX 3 Max follows H3 Max, fal Research’s previously released post-trained video model. fal said H3 Max produced video in 3.49 seconds in its benchmark, which it described as faster-than-real-time generation, while achieving top-tier image-to-video quality in independent evaluations. The product page notes that H3 Max, at 7.55 seconds, is the only faster model in its published comparison.

fal Head of Research Kaan Oktay said the team achieved a 50-point Elo improvement while making inference more than 11 times faster, and that fal Research is continuing to post-train the model. fal describes this launch as the first iteration of FLUX 3 Max, with further improvements to visual quality, generation speed, and overall performance underway.

Reported Quality and Speed Figures

fal reported that FLUX 3 Max reached a 1,139 Elo rating across 4,868 votes in preliminary independent human preference evaluations on Artificial Analysis Video Arena, a result the company said places the model among the highest-quality image-to-video systems evaluated. fal’s latency benchmark in the release measured generation of a 10-second video at 10.95 seconds, which the company described as 22 times faster than Seedance 2.5’s 277.80 seconds at the same reported Elo rating and 11.5 times faster than the original FLUX 3’s 126.44 seconds. The release also reports a 50-point image-to-video Elo gain over the original FLUX 3, from 1,089 to 1,139.

The product page publishes a different latency figure. It lists a 12.45-second average for a 10-second 720p video, describing the improvement over FLUX 3 as roughly 10 times faster, and a comparison table listing Gemini Omni Flash 1.1 at 1,134 Elo and 47.37 seconds, MiniMax H3 at 768p at 1,141 Elo and 83.26 seconds, Wan 3.0 at 1,127 Elo and 180.16 seconds, Seedance 2.5 at 1,139 Elo and 277.80 seconds, and the original FLUX 3 at 1,089 Elo and 126.44 seconds. The page identifies the Elo figures as Artificial Analysis’s preliminary AA-Video-I2V v1.1 ratings and the generation times as averages for a 10-second 16:9 video, and it describes FLUX 3 Max as tied for first place in image-to-video quality while noting that MiniMax H3 is listed first on that board.

Artificial Analysis’s image-to-video leaderboard, which ranks models by Elo derived from blind pairwise human preference votes cast by validated evaluators, listed FLUX 3 Max at 720p with a 1,138 Elo rating plus or minus 15 across 5,054 samples, ranked second behind MiniMax H3 at 768p with 1,140 Elo. The board lists Seedance 2.5 at 1,138 Elo and the original FLUX 3 at 1,086 Elo, and it dates FLUX 3 Max’s release to October 2026.

Availability and Launch Pricing

FLUX 3 Max is served through fal’s inference platform under a unified API. The announcement describes access to text-to-video, image-to-video, and keyframes-to-video generation, while the product page lists two endpoints at launch, text to video and image to video, both at 720p. According to the page, content generated through the fal.ai API may be used in commercial projects under fal’s terms.

Launch pricing is set at $0.0425 per second of 720p output video, billed per second with no minimums and no subscription. The product page states that the rate is 75 percent off the $0.17-per-second list price and runs until November 7, 2026, and fal’s release describes the discount as running for the first three weeks.