fal Launches FLUX 3 Max Video Model

Generative media company fal has launched FLUX 3 Max, a video generation model post-trained by its research team together with Black Forest Labs, the German lab behind the FLUX models. In a press release on 19 October 2026, fal said the model delivers “frontier visual quality at industry-leading inference speeds”. It builds on the FLUX 3 foundation model through specialised post-training, reinforcement learning and inference optimisation.

The headline numbers are fal’s own. The company says FLUX 3 Max generates a 10-second video in 10.95 seconds in its latency benchmarks. It also cites a preliminary Elo rating of 1,139 on Artificial Analysis Video Arena, an independent human-preference leaderboard. Both figures point to a wider shift: video model gains increasingly come from tuning and serving an existing model, not only from training a new one.

In brief

fal and Black Forest Labs have released FLUX 3 Max, a video model post-trained by fal Research on top of Black Forest Labs’ FLUX 3.
fal says it scored a preliminary 1,139 Elo across 4,868 votes on Artificial Analysis Video Arena, up from 1,089 for the original FLUX 3.
In fal’s benchmarks, it makes a 10-second video in 10.95 seconds, 11.5 times faster than FLUX 3; it is available through fal’s platform from 19 October.

What fal says FLUX 3 Max achieves

According to fal, “In preliminary independent human preference evaluations on Artificial Analysis Video Arena, FLUX 3 Max achieved a 1,139 Elo rating across 4,868 votes, placing it among the highest-quality image-to-video models evaluated.” The release adds that the result has “overlapping confidence intervals among the highest-ranked models”. In other words, it sits in a top group, not clearly alone at the top.

On speed, fal says the model generates “a 10-second video in 10.95 seconds, compared with 277.80 seconds for Seedance 2.5, at the same reported image-to-video Elo rating”, which it describes as over 22 times faster. Against the original FLUX 3, fal reports image-to-video Elo rising “from 1,089 to 1,139” and generation latency falling “from 126.44 seconds to 10.95 seconds”, an 11.5-fold speed-up. These latency figures come from fal’s own benchmarks.

Measure (fal’s figures)	FLUX 3	FLUX 3 Max
Image-to-video Elo (preliminary for Max)	1,089	1,139
Time to generate a 10-second video	126.44s	10.95s

How the collaboration works

fal says Black Forest Labs brought foundation-model expertise, while fal Research handled post-training, reinforcement learning, inference optimisation and systems engineering. Rather than simply deploying an existing model, fal says its team worked directly with FLUX 3 to improve visual fidelity, motion coherence and prompt adherence. The approach extends work behind H3 Max, fal’s earlier video model, which the company says produced video in 3.49 seconds in its benchmark.

“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.”

Burkay Gur, fal’s CEO, said: “We believe some of the most important advances in generative AI will come from optimizing models and the systems that run them together.” Kaan Oktay, fal’s head of research, added that “Model performance doesn’t stop improving when base model training ends.”

Availability and what comes next

fal says FLUX 3 Max is available from 19 October 2026 through its inference platform, offering text-to-video, image-to-video and keyframes-to-video generation through a unified API. The company calls this the first iteration of the model and says fal Research is continuing to post-train it, with further gains in quality and speed under way.

Our read

The Elo figure is preliminary, and the speed comparisons are fal’s own measurements, so independent tests will matter. But the claim that post-training and inference work alone can lift both quality and speed this much is significant. It suggests open foundation models such as FLUX 3 may increasingly be turned into competitive products by the platforms that serve them, not just by the labs that train them.