Black Forest Labs releases FLUX 3 Action, an open 7B world-action model for robots

Black Forest Labs has released FLUX 3 Action, an open-weight model for robotics built on its multimodal FLUX 3 model, which was trained mainly on video along with image and audio data. As a world-action model, it takes multi-camera video feeds from a robot's workspace and predicts both the next action an agent should take and how the environment will change as a result. BFL says FLUX 3 Action sets a record success rate on the RoboLab-120 leaderboard with seven billion parameters. That is less than half the size of the previous best open model, and BFL says it runs up to 3.95 times faster. The company argues that large reasoning models can plan well but are often too slow and too large for robots, so efficiency matters for running on the device. BFL also sees uses in digital settings, with video games as testing grounds for navigation and, later, fast-reacting agents that operate computers. The weights are on Hugging Face, so robotics researchers and companies can test the claims themselves. The benchmark results are the company's own. The excerpt does not give the license terms.