FLUX 3 Action: Open-Weights 7B World Action Model

Loading video

Loading video

Black Forest Labs released FLUX 3 Action on Hugging Face: an open-weights 7B world action model that takes camera frames, robot state and a text instruction, and returns the next chunk of actions denoised together with the next video frames. The base is an adaptation component rather than a complete policy - new embodiments train their own action heads - and ships with shared frozen encoders (video VAE plus a Qwen3-VL-4B-Instruct text encoder) used by the SO-101 and DROID policies via LeRobot.