FLUX 3 Action: An Open Robotics Model From Black Forest Labs

Black Forest Labs has released FLUX 3 Action, an open robotics model with 7 billion parameters. It is built on the multimodal FLUX 3, trained on video, images, and audio, and the weights are publicly available on Hugging Face. The model takes in video from multiple cameras at the robot's work site and predicts what action to take next and how the surroundings will change. On the RoboLab-120 benchmark, it took first place among open models at less than half the size of the previous leader. Speed is up almost fourfold, by a factor of 3.95 according to the company.

Here an open model with 7 billion parameters beats a model twice its size on a benchmark built for this exact task. That is the same trend I see in language models: a good architecture matters more than raw size. I do not work with robots myself. For agents I use Claude Code. But if a model like this fits directly on the robot instead of running only in the cloud, that removes the dependence on a constant internet connection. Black Forest Labs does not give exact numbers for inference cost, though. Right now this is just open weights without a finished product. How well it holds up outside the benchmark will become clear once someone builds a real robot on top of it.