FLUX 3 Action: Open Source Robotic Model By Black Forest Labs

Black Forest Labs released their latest open source robotic model, FLUX 3 Action. It contains 7 billion parameters and surpassed the state of the art on the RoboLab-120 benchmark by half its size! The FLUX 3 Action model is built upon the multimodal FLUX 3 model, which was trained on video, image, and audio data. You can access the model weights here on Hugging Face. When provided with a multi-view perspective of the robot’s environment, this model can predict future actions and changes within the environment. The FLUX 3 Action model placed first on the RoboLab-120 benchmark for open models, while being significantly smaller than the previous state of the art. They state that the speed is improved by a factor of 3.95.

It’s incredible to think that an open model with fewer parameters than another model was able to outperform it on a task tailored towards it. In my opinion, architecture matters more than raw parameters when it comes to language models too. While I don’t work with robots, I do utilize Claude Code for my agents. If these models are able to be deployed onto the robot itself instead of just the cloud, then we won’t require constant internet access. However, they don’t specify any details regarding the costs of running the model. This is an open weight model, so we’ll have to wait until someone implements this model into a physical robot to see how effective it is in practice.