Black Forest Labs Launches FLUX 3 Action, an Open 7B Robotics Model

Black Forest Labs Enters Physical AI with FLUX 3 Action

On September 24, 2026, Black Forest Labs (BFL) announced the release of FLUX 3 Action, an open AI model developed for robotics and spatial decision-making. The model weights have been released publicly on Hugging Face, marking an expansion for the company beyond generative media into embodied AI.

Multimodal Action Prediction in a 7-Billion-Parameter Frame

FLUX 3 Action operates as a world-action model. By ingesting multi-camera video feeds from a robot's active workspace, the system simultaneously determines what physical movement the agent should execute next and forecasts how the surrounding environment will alter as a consequence.

The architecture derives from BFL's foundation FLUX 3 multimodal model, which was trained primarily on video data alongside audio and still imagery. Packaged at 7 billion parameters, FLUX 3 Action captured a new success-rate record on the RoboLab-120 benchmark. According to BFL, the model achieves this while being less than half the parameter size of the previous record-holding open model and running up to 3.95 times faster.

Bridging the Gap Between Reasoning and On-Device Control

While contemporary frontier reasoning models demonstrate sophisticated high-level planning, their latency and compute footprints historically restrict real-world robotic integration. BFL positioned FLUX 3 Action’s smaller footprint and execution speed as direct solutions for the constraints of low-latency, on-device deployment.

Beyond physical robotics, BFL identified software-based applications for the system. The lab pointed to video games as practical sandboxes for evaluating agent navigation, noting that the underlying architecture could eventually support fast-reacting agents designed to control computer interfaces.