The Pulse
fal Launches FLUX 3 Max for Faster AI Video Iteration
fal says its FLUX 3 Max model can generate a 10-second video in 10.95 seconds, while scoring 1,139 Elo in preliminary human-preference evaluations. The model, developed with Black Forest Labs, is available through fal’s API for text-to-vide
AI.info Team · Written with AI · fact-checked ·
“We built FLUX 3 as a foundation for others to build on,” said Robin Rombach, co-founder and CEO of Black Forest Labs. “FLUX 3 Max is a great example of what that looks like. fal took our model and customized it for the developers and enterprises on their platform. That's the kind of collaboration FLUX was designed for, and we're excited to see it reach more of the people building with it.”
Robin Rombach, co-founder and CEO of Black Forest Labs
FLUX 3 Max cuts a 10-second render to 10.95 seconds
fal launched FLUX 3 Max on October 19, a video-generation model developed with Black Forest Labs that the companies say can return a 10-second video in 10.95 seconds. The release positions the model’s speed as a way to make repeated generation and revision less of a waiting game for developers building video tools.
The model starts from Black Forest Labs’ FLUX 3 foundation model. fal Research post-trained it and optimized inference, with the stated aim of improving visual fidelity, motion coherence and prompt adherence while reducing generation time. fal describes the launch as the first iteration, with additional post-training underway.
Rombach’s comment frames the release as an example of developers adapting a foundation model for specific uses, rather than treating the original model as a finished product. The work is a joint effort: Black Forest Labs supplied the base model, while fal says its research and systems teams handled post-training and inference optimization.
The benchmark compares quality as well as speed
In preliminary human-preference evaluations on Artificial Analysis Video Arena, FLUX 3 Max scored 1,139 Elo across 4,868 votes, according to fal. The company says the rating puts it among the highest-ranked image-to-video models, with overlapping confidence intervals among the top models—a qualification that means the evaluation does not establish a clear lead over every nearby competitor.
fal’s latency comparison puts FLUX 3 Max’s 10-second generation at 10.95 seconds, against 277.80 seconds for Seedance 2.5. The release says both models have the same reported Elo rating in that comparison, making the speed claim a comparison at a matched quality score rather than a claim that FLUX 3 Max produces a higher-rated result.
Against the original FLUX 3, fal reports a 50-point Elo increase, from 1,089 to 1,139, and a reduction in generation latency from 126.44 seconds to 10.95 seconds. The release characterizes that as 11.5 times faster inference. Those figures come from the company’s stated benchmark comparisons; the release does not provide enough methodological detail to treat them as a general performance guarantee for every prompt or production setup.
Three video workflows are available through fal
At launch, developers can use the model for text-to-video, image-to-video and keyframes-to-video generation through fal’s inference platform and a unified API. That gives teams options to start from a prompt, animate an image or supply keyframes to guide a clip’s progression.
fal says FLUX 3 Max is available at a 75% discount for its first three weeks. The launch announcement does not state a per-second price, so the discount alone does not show what a particular workload will cost.
Fast generation changes the revision loop
A 10-second clip generated in roughly 11 seconds could let a developer test a prompt, inspect the result and adjust it without waiting several minutes for each attempt. That can matter in workflows where teams compare motion, framing or prompt changes while building a creative application. The benchmark supports a claim about reported generation latency, not a guarantee that full production workflows—including queueing, review and revisions—finish in the same interval.
fal says its researchers are continuing to post-train the model, so the launch figures describe an initial version rather than a final performance target. For now, the clearest concrete offer is access through fal’s API to three video-generation modes, with a 10-second clip reported at 10.95 seconds in the company’s latency benchmark.