Black Forest Labs Debuts FLUX 3 Action Robotics Model

Black Forest Labs has officially released FLUX 3 Action, an open-source artificial intelligence model tailored specifically for robotics applications. Built upon the company's multimodal FLUX 3 foundation model—which underwent training on video, image, and audio data—this new release operates as a world-action model. By processing multi-camera video feeds directly from a robot's workspace, FLUX 3 Action can predict both the next physical action an agent should execute and how those movements will alter the surrounding environment.

The model achieves these capabilities while maintaining a highly compact footprint. Featuring just seven billion parameters, FLUX 3 Action has set a new success rate record on the RoboLab-120 leaderboard. According to Black Forest Labs, the model is less than half the size of the previous record-holder in the open-source category. Furthermore, it operates up to 3.95 times faster than its predecessor, a speed increase that addresses a major bottleneck in robotic control.

For robotics practitioners and developers, this release represents a significant shift toward viable on-device deployment. While massive reasoning models excel at high-level planning, their computational weight and latency make them impractical for real-time physical control. FLUX 3 Action offers a lightweight alternative that can run locally on robot hardware without sacrificing performance. Developers can already access the model weights, which have been made publicly available on Hugging Face.

Beyond physical robotics, Black Forest Labs envisions broader applications for the model in digital environments. The architecture could be utilized within video games as a testing ground for navigation systems. Eventually, the company suggests the technology could power fast-reacting digital agents capable of navigating and operating computer interfaces.