black-forest-labs/flux-3-action-base

FLUX 3 Action Base is an open‑weights 7 B world‑action model that ingests camera frames, the robot’s state and a textual instruction. It produces a chunk of predicted actions together with denoised video frames of the upcoming scene. The repository supplies the base weights and shared frozen encoders that were used in the SO‑101 and DROID policies, but it does not contain a complete control policy. A developer would load the flux‑3‑action‑base.safetensors file and attach the Qwen3‑VL‑4B‑Instruct text encoder from the textencoder directory. After that, new action heads must be fine‑tuned for the specific robot, simulator or game embodiment. The trained heads can then be combined with the shared encoders to generate motor commands and optional future video predictions. Safety limits such as joint velocity, force and workspace boundaries are not enforced by the model and must be imposed by the surrounding application. Validation should be performed in simulation or with the arm’s safety stops engaged before any deployment near people. The model’s license and the FLUX Kommunity License v.1.0 impose restrictions against illegal use, unsupervised operation that endangers humans, and abusive or harmful applications.