Black Forest Labs Releases FLUX 3 Action Robot Control Model
A 7 billion parameter model beating a 12 billion parameter one on Robolab-120 suggests that scale is not the only path to stronger robot control, and Black Forest Labs is now shipping it jointly with NVIDIA into Hugging Face's LeRobot and Jetson.
Reporting from 1 source: GIGAZINE.
Black Forest Labs released FLUX 3 Action, a 7 billion parameter world action model derived from its FLUX 3 multimodal AI model, for robot control. It scored 42.92% on the Robolab-120 benchmark. That is higher than the closed model OASIS WAMNVIDIA at 39.0% and NVIDIA's Cosmos3-Nano at 36.8%. Black Forest Labs says FLUX 3 Action also runs up to 3.95 times faster than Cosmos3-Nano.
Black Forest Labs frames FLUX 3 Action as a world action model: a video generation model predicts how the real world changes when an action is taken, and an inverse dynamics model turns that prediction into movement. The pitch is that simulating the future before acting cuts the cost of trial and error compared with robots that respond step by step from camera input.
Black Forest Labs said it integrated FLUX 3 Action natively into Hugging Face's open-source robot platform LeRobot jointly with NVIDIA, with fine-tuning recipes and edge deployment on NVIDIA Jetson. It also said the model produced promising results outside robotics, in simulation environments where it must read a visual environment and pick its next action.
Synthesized by Yomimono from the 1 cited source below, including Japanese-language reporting where cited, then editorially reviewed before publishing.