Black Forest Labs has released FLUX 3 Action, a 7-billion-parameter open-weights World Action Model for robot control that ranks first on the RoboLab-120 leaderboard with a 42.92% task success rate2. The model is less than half the size of the previous best open model and runs up to 3.95 times faster, according to Black Forest Labs3.
FLUX 3 Action reads camera frames, robot state, and a text instruction, then jointly predicts future video frames and the next chunk of robot actions. It builds on the multimodal FLUX 3 backbone, which was trained primarily on video but also on image and audio data, with video accounting for over 95% of training tokens.
Benchmark and evaluation results
On RoboLab-120, a simulated benchmark comprising 120 tabletop tasks in Isaac Sim with 10 trials each on a DROID-style Franka setup, FLUX 3 Action's multi-seed mean for the guidance-distilled FP8 checkpoint is 42.24% ± 0.36. That compares with 36.8% for NVIDIA's Cosmos 3 Nano and 28.0% for π0.5 on the same leaderboard.
Positronic Robotics ran a blind physical evaluation on a Franka arm across 10 DROID tasks with 3 attempts each. FLUX 3 Action completed 28 of 30 attempts, Cosmos 3 Nano scored 27 of 30, DreamZero scored 20 of 30, and π0.5 scored 13 of 30.
Speed and deployment trade-offs
Black Forest Labs measured that Cosmos 3 Nano in FP8 needs roughly 4.7 times more processing time than π0.5 in BF16 per second of robot motion on a B200. FLUX 3 Action closes that gap through a smaller backbone and distillation: the base and guidance-distilled checkpoints run 1.52 to 3.95 times faster than Cosmos 3 Nano in FP8 across consumer, workstation, and datacenter GPUs. The guidance-distilled checkpoint drops a second guidance pass, runs 1.8 to 2 times faster than the base policy, and scores 0.6 to 1.08 percentage points higher. A step-distilled checkpoint uses a single sampling step, running 3.15 to 4 times faster but with a 3.51 to 4.32 percentage-point drop in success.
Each inference call yields 32 actions at 15 Hz, covering 2.13 seconds of motion, compared with 1.0 second per call for π0.5. The DROID policy requires about 32 GB of GPU memory in BF16 on an H200 but fits on 24 GB cards with FP8 quantization and text encoder offload.
Licensing and integrations
The weights are available on Hugging Face under the FLUX Kommunity License, which allows non-commercial use. Black Forest Labs integrated the model into Hugging Face LeRobot with NVIDIA and supports edge deployment on NVIDIA Jetson. The lab also published a DROID recipe and an SO-101 LoRA recipe, with documentation showing an SO-101 pick-and-place skill learned from roughly 200 demonstrations.
Black Forest Labs tested hybrid control pairing FLUX 3 Action with GPT 6 Astra: at low reasoning effort, the hybrid solved 90% of episodes at $8.77 and 8 minutes 8 seconds per success, while pure Astra at maximum effort solved 100% but cost $13.47 and 16 minutes 23 seconds per success.
ANALYSIS By matching or exceeding larger world-action models on both accuracy and inference speed, FLUX 3 Action positions Black Forest Labs, previously known for image generation, as a competitor in the open-weights robotics-model space.