Black Forest Labs has released FLUX 3 Action, a 7B open-weights World Action Model (WAM) designed for robot control. The model processes camera frames, robot state, and text instructions to concurrently predict future video frames and robot actions. It achieved a first-place ranking on the RoboLab-120 leaderboard with a 42.92% task success rate, outperforming NVIDIA’s Cosmos 3 Nano and other competitor models.

Derived from the multimodal FLUX 3 backbone, the model was pretrained primarily on video data before undergoing midtraining with action-aligned video across 14 embodiments. BFL addresses the high computational cost typically associated with video-predicting WAMs by utilizing a smaller backbone and distillation, allowing it to achieve faster motion prediction speeds than competing models.

Deployable under the non-commercial FLUX Kommunity License, the DROID policy runs on H200, 24 GB consumer GPUs via FP8 quantization, and workstation hardware. Independent real-hardware testing by Positronic Robotics on a Franka arm validated these simulation results, with FLUX 3 Action completing 28 out of 30 physical task attempts.

Why it matters

  • Open-weights release provides robotics startups with a state-of-the-art 7B world action model that outpaces proprietary benchmarks.

  • Distillation and quantization allow real-time 15 Hz inference on 24 GB consumer and workstation GPUs, lowering hardware deployment barriers.

  • Demonstrates that pretraining on mixed video data dramatically improves task success rates compared to training solely on domain-specific robot demonstrations.

Source: marktechpost.com