Black Forest Labs posts open robotics model, says it beats larger rivals
The 7-billion-parameter FLUX 3 Action takes camera frames and a text instruction and returns a robot's next actions, under a licence the company wrote itself.
Black Forest Labs released FLUX 3 Action on Thursday, a 7-billion-parameter model for robots that the company says sets record success rates on the RoboLab-120 leaderboard. The weights are on Hugging Face under the FLUX Kommunity License v.1.0, and the collection carries two policies and a shared base.
The model card describes what the model does in plain terms. It takes camera frames, the robot's state and a text instruction, and returns the next chunk of actions. The SO-101 policy has 34 likes and 206 downloads in the past month; the DROID policy has 21 likes and 121 downloads.
The performance claims are the company's own. Black Forest Labs said the model is less than half the size of the previous best open model on RoboLab-120 and runs up to 3.95 times faster, according to The Decoder's report of the launch. No independent test of either figure has been published yet.
The setup in the model card is narrower than the leaderboard claim may suggest. The SO-101 policy reads two cameras, one on the scene and one on the wrist, and takes six state and action dimensions. It predicts 42 actions, executes 32 of them at 30 Hz, then replans.
Black Forest Labs is known for image generation, and this is its first robotics release. The company said large reasoning models plan well but robotics needs efficiency for work on the device itself, and pointed to digital uses such as testing navigation and computer-control agents.