Black Forest Labs' FLUX 3 Action sets RoboLab-120 record

Black Forest Labs is best known for the FLUX family of image models, and FLUX 3 Action extends that line into robotics. The new model builds on FLUX 3, a multimodal system trained mainly on video but also on image and audio data, and reframes it as what BFL calls a world-action model — one that reads camera feeds from a workspace and forecasts both the next action for an agent and how the scene will change.

The economics of the announcement rest on size. BFL says FLUX 3 Action reaches a record success rate on the RoboLab-120 leaderboard with seven billion parameters, less than half the size of the previous best open model, while running up to 3.95 times faster. In BFL's telling, large reasoning models plan well but are often too slow and too bulky to sit inside a robot, which is why this kind of efficiency is critical for on-device deployment.

Whether that translates into practical advantage hinges on how the model performs beyond RoboLab-120 and on real hardware, since the record is BFL's own characterization of a single benchmark. BFL also flags digital environments as an early proving ground, with video games serving as testing grounds for navigation and fast-reacting agents that operate computers possibly following later — a direction the company presents as potential rather than committed.