AINEWS
The German lab pitches a single AI system for both video creation and robot control, but the competitive and deployment claims rest on the company and its partners, not independent benchmarks.
A small German AI lab called Black Forest Labs released a new video generator this week called FLUX 3, and the company's pitch is unusual: the same system that produces short videos from text prompts is being tested as the brain inside factory robots.
FLUX 3 generates videos, images, and synchronized audio from a single architecture, according to the company's release post. The output can carry dialogue in multiple languages, animate a still image, continue a video from a final frame, and chain short clips into longer sequences with a text instruction. Those capabilities put it in the same category as Google's Gemini Omni video system, xAI's Grok Imagine, and ByteDance's Seedance 2.0.
Latent Space's AINews coverage of the release uses hedged language, describing the result as "perhaps SOTA" (state of the art, the best publicly reported score on a standard benchmark) rather than a confirmed benchmark win. The "beats Seedance 2.0, Gemini Omni, and Grok Imagine" framing traces to BFL's own positioning and a BFL-aligned newsletter, not to a public, reproducible head-to-head evaluation.
The same architecture now has a second job. BFL co-developed a variant called FLUX 3 x mimic with mimic robotics, a German robotics startup. The system takes a video and predicts what action a robot should take next: gripping, lifting, placing. The same model that generates pixels also generates motion commands. BFL says the system is already running on robots at an Audi factory.
The deployment claim is partner-sourced. The Audi robots are described in the BFL release. There is no independent verification of how many robots are running, what tasks they perform, or whether the work is a controlled demo or a production line. "Deployed" in a factory context can mean a one-week pilot, and BFL's post does not specify.
The robotics claim rests on a separate paper. A December 2025 arxiv preprint from the mimic team introduces a "Video-Action Model" (VAM) and argues that current robot-control systems are "blind to physical causality" because they pretrain on disconnected static images and text scraped from the web. A VAM trained on video, the paper argues, already knows that a falling object will hit the floor, which is what a robot needs to know.
The bet behind both products is that video generation and robot control are the same problem. Teaching a model to predict the next frame of a video also teaches it how objects hold together, how things move through space, and how events sound. That is the definition BFL gives of a "world model": a system that has learned a representation of how the world works, not just how it looks. BFL shipped the original FLUX 1 image model in 2024, so the FLUX 3 release two years later is, on the company's own telling, the completion of a roadmap toward a single learned representation of the world.
The competitive frame is BFL's, not the public's. The "beats Google, xAI, and ByteDance" framing is the lab's own positioning. An open-weights Dev release (one where the model's internal parameters are publicly downloadable so outside researchers can run their own tests rather than trust the lab's numbers) is reportedly still on the way, per Latent Space, which would let third parties run their own comparisons. Until then, the "perhaps SOTA" hedge is the most accurate available read.
The first public test will be the FLUX.3 Dev release. Once the model is downloadable, outside researchers can run the same video-quality comparisons BFL has not yet published, and try bolting the architecture to a robot arm to see whether the physics the model learned from video translates to motion commands a factory will trust.