FLUX 3 Action

Not yet independently verified. Black Forest Labs’ own model page dates the release only as “September 2026”. The 23 September date rests on VentureBeat’s report of that day; the Hugging Face repository was created on 22 September. Both benchmark results are the developer’s own. We will update this when it can be confirmed, and remove this note.

FLUX 3 Action is a 7-billion-parameter “world action model” from Black Forest Labs, published with open weights on Hugging Face and reported by VentureBeat on 23 September 2026. It adapts the multimodal backbone of the company’s FLUX 3 image and video model to robot control: given camera images, the robot’s joint positions and a plain-language instruction, it predicts a short sequence of motor commands together with frames showing what it expects to happen next. Black Forest Labs has published a shared base and two fine-tuned policies — one for the Franka arm, trained on the DROID dataset, and one for the low-cost SO-101 arm used with the LeRobot library — along with fine-tuning code. The Franka policy needs about 32 GB of GPU memory in BF16. The company reports 42.92 per cent task success on NVIDIA’s RoboLab-120 simulation benchmark (its own page also shows 42.2 per cent for a distilled variant) and 28 of 30 tasks completed in a real-robot test run by a partner; both are the developer’s own results and are not stored as scores. The licence is the thing to read before building on it: the FLUX Kommunity License v1.0 allows commercial use only for organisations with under US$5 million in annual revenue, so a larger company needs a separate agreement, and military, surveillance and biometric uses are prohibited. No hosted API or price had been announced.