Black Forest Labs' FLUX 3 now runs robots
Black Forest Labs has previewed FLUX 3, a jointly trained image, video and audio model whose backbone also powers FLUX-mimic, a robot-control model the company says is deployed at Audi.
Black Forest Labs (BFL) revealed an early version of FLUX 3, its next foundation model. Unlike FLUX 1 and FLUX 2, which generate images, FLUX 3 is trained jointly on images, video and audio from the start. BFL gave early access to mimic robotics, and the two companies built FLUX-mimic, a video-action model that the companies say has been tested and deployed on robots at Audi.
BFL's argument is that a model has to learn contact, motion, weight and cause and effect to generate realistic video, so robot actions are "one more view" of the same physical world. Video prediction accounts for more than 95% of FLUX 3's training compute. Audio makes up less than 0.5% of the tokens in a 720p video with audio.
To test this, BFL added action prediction partway through a large training run. Human ratings of text-to-video and image-to-video output first fell by up to 10%. After 3,500 steps, video quality was back to its earlier level, and the model could also predict actions. BFL says this shows one backbone can handle both tasks.
FLUX-mimic follows an approach first used in mimic-video. A lightweight action decoder reads intermediate features from FLUX 3's video prediction path. The announcement also cites backbone latency under 80ms and a full-system reaction time of 101ms.
All of these figures come from the two companies and have not been independently checked. BFL has not given a date for a full public release of FLUX 3.