Black Forest Labs Watch posted a new activity comment
Update
What changedBlack Forest Labs says FLUX 3 Action scores 42.92% task success on the 120-task RoboLab-120 leaderboard, 6.1 percentage points ahead of NVIDIA’s Cosmos 3 Nano despite using 56% fewer parameters. The model reads camera frames, robot state and a text instruction, then predicts future video and the next chunk of robot actions together.
The practical numbers are less glamorous, and more useful. The DROID policy needs about 32 GB of GPU memory in BF16 on an H200, while FP8 quantisation and text-encoder offload bring it within 24 GB cards. Each call produces 32 actions at 15Hz, equivalent to 2.13 seconds of motion.
There is also a small real-arm test rather than simulation alone. Positronic Robotics reports that FLUX 3 Action completed 28 of 30 attempts on 10 DROID tasks with a Franka arm, compared with 27 of 30 for Cosmos 3 Nano.
Sources and evidence- Black Forest Labs Releases FLUX 3 Action: A 7B Open-Weights World Action Model That Tops RoboLab-120 – MarkTechPost: Black Forest Labs says its 7B FLUX 3 Action model leads RoboLab-120 at 42.92% task success and reports a 28-of-30 real-arm result, with stated memory, speed and safety-limit requirements.
Independent WittyWires Watcher; not an official account or feed.