FLUX 3 learns about the world through image, sound, and action

Black Forest Labs' FLUX 3 model generates twenty-second clips with native audio. A multimodal approach that extends all the way to robotics.

Known for its image generators, Black Forest Labs is expanding FLUX into a single multimodal model, trained end-to-end on image, video, and audio, and capable of predicting actions. The starting hypothesis can be summed up in one sentence: each modality is only a partial capture of the same reality, and learning them together forces them to mutually constrain one another, with sound having to match impact and motion having to obey mass. The laboratory places FLUX 3 within its work on World Models.

The video component is opening in early access. Clips reach twenty seconds with natively produced sound in a single pass, generated from text, an initial image, a reference video, or keyframes used to frame transitions. Multilingual dialogues are supported, and multiple shots chain together into longer sequences where visual references maintain the same character from one scene to another. Black Forest Labs describes its evaluations as preliminary and claims preference rates ahead of several competing video models. The image component will follow.

The robotics side stems from the same hypothesis. According to the laboratory, video prediction absorbs more than 95% of the computational cost of training, because rendering a scene believably requires integrating contact, weight, and causality. Action then falls under the same register, a brief description of the robot's state closely linked to what it observes. With the company mimic robotics, a lightweight action decoder was grafted onto the internal representations of the video backbone, resulting in FLUX-mimic, which was tested on Audi production lines for tasks that had remained manual, such as assembling parts bins, inserting ECUs into tight slots, and handling seals and flexible cables. The backbone transitions from the image to its representation of the world in less than 80 milliseconds on a single RTX 5090 card, with the complete system reacting in 101 milliseconds.

The rollout will take place in stages, with video and audio delivered via an API and private access to the weights, action prediction remaining reserved for research and industrial partners, followed by image, and ultimately, a multimodal backbone released in open weights under the name FLUX 3 Dev.