A company known for image generators releases a model for robots. It sounds like a turn, but it is a natural continuation.
If a model can predict the next frames of a video, it already understands a little about how things move. Black Forest Labs take that understanding and make it predict the arm's next move as well: what the robot will do and how the scene will look afterwards.
Fact
The facts: on 22 September 2026 Black Forest Labs published FLUX 3 Action - an open-weight model with 7 billion parameters, derived from their multimodal FLUX 3 backbone, trained on video, images and audio, with video making up over 95% of the data. The model jointly predicts future video frames and robot actions from camera frames, joint positions and a text instruction. Per the company, the single-step variant reaches 38.3% success on the RoboLab-120 ranking and the guidance-distilled variant 42.2%, against 36.8% for NVIDIA's Cosmos 3 Nano with 16 billion parameters and 28.0% for Pi0.5, and in quantised form it is 1.52 to 3.95 times faster than Cosmos 3 Nano. Ready variants for the SO-101 and DROID robots are also published. The weights are under the FLUX Kommunity License v1.0, which allows use for non-commercial purposes, including non-commercial robotics; other commercial use requires a separate licence from the company, with an exception for firms with annual revenue under 5 million dollars, which may use the outputs commercially under conditions.
Fast and smart together
The most interesting experiment in the report is not the ranking. BFL put a large reasoning model, GPT-6 Astra, in charge, with FLUX 3 Action doing the work and the boss stepping in only when needed.
Pure reasoning solves every task, but by their figures each success costs $13.47 and about 16 minutes. The combination solves 90% and costs $8.77 for about 8 minutes. The slow thinker with fast hands comes out cheaper than the thinker doing everything alone.
A good foreman does not move every brick himself. He knows when to step in.
The ranking comes from simulation, and both it and the costs are the company's. There is also a test on a real arm: the independent Positronic Robotics ran the model on a Franka robot in its lab, and it completed 28 of 30 attempts across ten tasks. Thirty attempts are a start, not a verdict.
The weights are free for research and hobby. For a robot that will work at a client's, the licence is read first, because robotics is named in it explicitly.
Official primary sources
Black Forest Labs - FLUX 3 Action: A 7B World Action Model for Robot Control, 22.09.2026Hugging Face - black-forest-labs/flux-3-action-base (model card and LICENSE)The visual is generated code-art, no third-party images.