Black Forest Labs releases 7B FLUX 3 Action model

Black Forest Labs released FLUX 3 Action, a 7B open robotics model leading RoboLab-120 at 42.92%, with sources differing on the 3.95x speed claim.

Black Forest Labs released FLUX 3 Action, a 7-billion-parameter open-weights world-action model for robotics built on multimodal FLUX 3, which The Decoder says was trained mainly on video plus image and audio. The Decoder reports that it takes multi-camera workspace video and predicts the agent’s next action and how the environment will change; MarkTechPost says it jointly predicts future video frames and robot actions from camera frames, robot state, and text. On RoboLab-120, MarkTechPost reports 42.92% task success, ahead of NVIDIA’s 16B Cosmos 3 Nano at 36.8% and pi0.5 at 28.0%—a 6.1-point lead with fewer parameters—while The Decoder says the model set a success-rate record and is under half the size of the previous best open model. On speed the sources differ: The Decoder says it runs up to 3.95 times faster, whereas MarkTechPost specifies that FP8 checkpoints run up to 3.95 times faster than Cosmos 3 Nano. A blind Franka evaluation completed 28 of 30 attempts. Weights are on Hugging Face under the non-commercial FLUX Kommunity License, with BF16 and FP8 checkpoints and LeRobot and Jetson support; The Decoder also cites games and computer-use agents as later applications.

• Black Forest Labs released FLUX 3 Action, a 7B open-weights world-action robotics model built on multimodal FLUX 3 (trained mainly on video plus image and audio, per The Decoder).
• The Decoder: it takes multi-camera workspace video and predicts the next action and scene change; MarkTechPost: it jointly predicts future video frames and robot actions from camera frames, robot state, and text.
• MarkTechPost: RoboLab-120 task success is 42.92%, ahead of NVIDIA Cosmos 3 Nano at 36.8% (16B) by 6.1 points and pi0.5 at 28.0%.
• The Decoder says it set a RoboLab-120 record and is under half the size of the previous best open model.
• Speed claims differ: The Decoder says up to 3.95x faster; MarkTechPost says FP8 checkpoints run up to 3.95x faster than Cosmos 3 Nano.
• A blind Franka evaluation completed 28 of 30 attempts (MarkTechPost).
• Weights are on Hugging Face under the non-commercial FLUX Kommunity License, with BF16 and FP8 checkpoints plus LeRobot and Jetson support.
• The Decoder cites games and computer-use agents as later applications.