Black Forest Labs launches FLUX 3 Action, an open robotics AI model

Black Forest Labs has released FLUX 3 Action, an open-weight AI model built for robot control. The weights went up on Hugging Face on September 24, 2026, together with a technical report that documents how the model was trained and fine-tuned.

A world-action model on the FLUX 3 backbone

FLUX 3 Action derives from the multimodal FLUX 3 model, which was pretrained on a large collection of video, image and audio data with a strong emphasis on video. It is what BFL calls a world-action model: it takes multi-camera video feeds from a robot workspace and the robot's joint positions, then predicts both the next action the agent should take and how the environment will change as a result. The robot executes a set number of predicted actions, looks at the scene again, and replans from the fresh observation.

A record at half the size

On the RoboLab-120 leaderboard, the single-step 7-billion-parameter checkpoint reaches a 38.3% success rate, ahead of the 36.8% scored by Cosmos 3 Nano, the previous best open model, at less than half its parameter count. A guidance-distilled variant lifts the record to 42.2%. BFL says both checkpoints run up to 3.95 times faster than Cosmos 3 Nano in FP8 on consumer, workstation and datacenter GPUs, and that the model predicts a horizon of 2.13 seconds of robot motion against one second for the Pi0.5 policy.

Why efficiency decides the case

Frontier reasoning models plan well, BFL argues, but they are often too slow and too bulky for robots, which makes efficiency critical for on-device deployment. The report puts numbers to that claim: FLUX 3 Action spends about 21 milliseconds of compute per second of robot motion and costs roughly 9 cents per successful rollout on an H200 GPU, while a reasoning-only approach needs about 36 seconds of compute per second of motion and $13.47 per success. Combining a fast policy with a reasoning model that steps in only when needed solves 90% of episodes and brings the cost down to $8.77 and eight minutes per success — up to 53.64% more success per dollar than the best alternative setup.

Beyond the robot arm

BFL also sees potential in digital environments. It uses video games as a testbed for navigation, framing fast-reacting agents as a step towards computer-use systems that respond to visual input.