FLUX 3 Action Hits Record on RoboLab-120 with 7 Billion Parameters

Black Forest Labs releases the weights of FLUX 3 Action on Hugging Face. The action model for robotics, derived from FLUX 3 multimodal, claims a record on RoboLab-120 with seven billion parameters and execution speeds up to 3.95 times faster than the previous best open model, according to the company.

Available Weights and Advanced Digital Use Cases

The weights files for FLUX 3 Action are accessible via Hugging Face. Black Forest Labs notes that this model is of interest for applications in digital environments, citing video games as experimental platforms for navigation or for reactive agents capable, ultimately, of taking control of computers. According to the company, large reasoning models are effective for planning but often prove too slow and bulky for robotic use, making high efficiency necessary to operate directly on-device.

Multi-Camera Inputs and Predictions of Actions and Effects

FLUX 3 Action processes multi-camera video streams from a robotic workspace. It predicts the next action an agent should take as well as how the environment will change afterward. This model is based on FLUX 3, a multimodal model primarily trained on video data, but also on images and audio.

Claimed Record on RoboLab-120 with 7 Billion Parameters

Black Forest Labs states that FLUX 3 Action achieves a record success rate on the RoboLab-120 ranking. This model includes seven billion parameters, which is less than half the size of the previous best open-source model, and it reportedly displays speeds that can be up to 3.95 times greater.