Black Forest Labs has released FLUX 3 Action, a 7B parameter open-weights World Action Model (WAM) for robot control that ranks first on the RoboLab-120 leaderboard with a 42.92% task success rate.

  • The model predicts future video frames and robot actions jointly by reading camera frames, robot state, and text instructions.
  • It outperforms NVIDIA's Cosmos 3 Nano by 6.1 percentage points while using 56% fewer parameters.
  • BFL provides three checkpoint recipes: Base, Guidance-distilled (1.8x to 2x faster), and Step-distilled (up to 4x faster).
  • The model supports fine-tuning via LoRA and is integrated into Hugging Face LeRobot for edge deployment on NVIDIA Jetson.
  • Real hardware evaluations on a Franka arm showed FLUX 3 Action completing 93.3% of attempts, surpassing Cosmos 3 Nano's 90%.

The release offers a deployable solution that balances the high accuracy of video-predictive models with the speed of vision-language-action models.