Black Forest Labs has released FLUX 3 Action, an open-weight model intended to turn video from a robot’s workspace into the next physical action. The seven-billion-parameter system receives multiple camera views, predicts what the robot should do and models how the scene is likely to change afterward.
The model extends FLUX 3, which was trained primarily on video with additional image and audio data. Black Forest Labs describes the new version as a world-action model: instead of only interpreting a frame, it links visual changes to control decisions. That design targets robots that must react quickly without sending every step to a very large remote model.
According to the company, FLUX 3 Action achieved the highest success rate on the RoboLab-120 leaderboard with less than half the parameters of the previous leading open model. It also ran as much as 3.95 times faster. Those are vendor-reported benchmark results and need independent testing across different robots, tasks and hardware.
The weights are available on Hugging Face, enabling researchers to inspect and adapt the system. Black Forest Labs also sees possible uses in games and computer-operating agents, where predicting the next state matters. Availability does not by itself solve deployment: teams still need hardware integration, safety limits and validation in the specific physical environment.