Black Forest Labs launches FLUX 3 Action, outperforming rivals with 7 billion parameters.

Black Forest Labs unveiled FLUX 3 Action on Monday in Berlin, a 7‑billion‑parameter “World Action Model” that translates camera feeds, robot state data and natural‑language commands into real‑world motions, and it already tops the most demanding robotics leaderboards despite being roughly half the size of competing systems.

The launch follows a rapid expansion of the German startup’s FLUX family, which first gained attention for high‑fidelity image and video generation. As industry players race to embed large‑scale AI into physical devices, the need for models that can run on modest compute while still handling complex perception‑action loops has become urgent. Earlier this year, OpenAI’s RT‑1 and DeepMind’s Gato set performance records but required 15‑20 billion parameters and heavyweight GPU clusters, limiting their use outside research labs. Black Forest Labs’ shift into robotics reflects a broader trend of democratizing embodied AI, aiming to bring sophisticated control to smaller manufacturers and research groups that lack deep‑learning infrastructure.

In benchmark tests on the RoboSuite and Meta‑World suites, FLUX 3 Action achieved a 92 percent success rate on pick‑and‑place, stacking and door‑opening tasks, beating the previous best by 4 points. The model’s architecture blends a vision transformer with a recurrent state encoder, allowing it to process continuous streams of sensory data without the latency spikes typical of larger networks. Because the weights are released under an open‑source licence, developers can fine‑tune the system for specific hardware, from collaborative arms in factories to low‑cost service bots. The announcement sparked praise from the robotics community; the European Association for Robotics called it “a watershed moment for accessible embodied AI,” while investors led by Earlybird Capital noted the potential to accelerate product cycles for startups that previously relied on proprietary, closed‑source solutions.

Black Forest Labs says it will publish the full training code and a lightweight inference library within the next month, and it is already in talks with several European robot manufacturers to integrate FLUX 3 Action into upcoming product lines. The company also plans a series of open challenges to crowdsource task‑specific adaptations, hoping to build a vibrant ecosystem around the model. If the early results hold up, the technology could lower the barrier for companies to add intelligent manipulation to their devices, speeding up automation across sectors ranging from logistics to healthcare.