Black Forest Labs Debuts FLUX 3 Action Robot Model

Black Forest Labs has launched FLUX 3 Action, a 7-billion-parameter open-weight World Action Model (WAM) for robot control. The model predicts a robot's next actions alongside future video frames in a single forward pass, projecting a 2.13-second action horizon. By utilizing a smaller backbone and distillation, it runs 1.45 to 3.95 times faster than competing WAMs while using 56 percent fewer parameters than the previous best open model.

On the RoboLab benchmark, the model's guidance-distilled checkpoint set a new state-of-the-art success rate of 42.2 percent, with a throughput 2.85 to 3.15 times higher than Cosmos 3 Nano. Its single-step checkpoint achieved a 38.3 percent success rate, outperforming Cosmos 3 Nano's 36.8 percent. Additionally, FLUX 3 Action processed motion 1.45 to 1.66 times faster than pi0.5. In physical testing on a Franka arm by Positronic Robotics, FLUX 3 Action succeeded in 28 of 30 attempts, beating Cosmos 3, DreamZero, and pi0.5.

The architecture reuses the FLUX 3 backbone, outputting 32 actions and predicted frames. To optimize deployment, step distillation compresses four sampling iterations into one, while guidance distillation removes the second forward pass of classifier-free guidance. Running FP8 with torch.compile yields a 1.4-times speedup on local GPUs like the RTX 5090. The model maps inputs to a 64-dimensional gaming space or a 50-dimensional EE50 pose space. During fine-tuning on the DROID dataset, the learning rate warms to 2e-4 over 2,000 steps after being frozen for 1,000 steps.

For practitioners, this lowers the computational barrier to real-time robotics. The model is on Hugging Face with LeRobot and NVIDIA Jetson support across three checkpoints: flux-3-action-base, flux-3-action-so101, and flux-3-action-droid. When paired with a GPT 6 Astra reasoning model, a hybrid setup solved 90 percent of tasks at $8.77 per success, which is 29 percent cheaper and 40 percent faster than using pure reasoning.