Black Forest Labs has released FLUX 3 Action, an open source model designed specifically for robotics applications. This new tool builds upon the existing FLUX 3 architecture, which was originally trained primarily on video alongside image and audio datasets. The system functions as a world-action model by processing multi-camera feeds from a robot workspace to predict the next required movement and the resulting environmental changes. On the RoboLab-120 leaderboard, the model established a new record for success rate using just seven billion parameters. This figure represents less than half the size of the previous best open model while running up to 3.95 times faster than its predecessors.
Large reasoning models often plan effectively but remain too slow and bulky for physical robots, making this specific efficiency critical for on-device deployment. The developers note that while large models handle complex planning well, their computational demands usually prevent real-time operation in hardware. FLUX 3 Action addresses this bottleneck by prioritising speed without sacrificing accuracy in dynamic environments. The weights are available for download on Hugging Face for researchers and developers to test. The team also sees potential applications in digital environments where video games could serve as testing grounds for navigation and fast-reacting agents that operate computers could follow down the line.
* Available for download on Hugging Face
* Runs up to 3.95 times faster than prior open models
* Uses fewer than half the parameters of the previous best
