Black Forest Labs' FLUX 3 Action tops RoboLab-120 with a 7B robot model

Black Forest Labs, the lab behind the FLUX image models, has released FLUX 3 Action, a 7-billion-parameter open-weights model for controlling robots. On the RoboLab-120 leaderboard, a benchmark of 120 tabletop tasks run in Isaac Sim simulation, it ranks first with a 42.92% task success rate.

The model takes in camera frames, the robot’s current state and a text instruction, then predicts both future video frames and the robot’s next chunk of actions at the same time. Black Forest Labs calls this a World Action Model, a category that pairs video prediction with motor control. That combination usually costs speed: rival World Action Models like Nvidia’s Cosmos 3 Nano lead RoboLab at 36.8% but are expensive to run, while faster action-only models called VLAs, such as π0.5, reach only 28.0%. Black Forest Labs says FLUX 3 Action closes that gap with a smaller backbone and a distillation process that trims sampling steps, rather than by giving up the joint video-and-action approach.

The company reports that, on a B200 chip, its base and guidance-distilled checkpoints run 1.52 to 3.95 times faster than Cosmos 3 Nano in FP8, an 8-bit numerical format that cuts memory and compute needs. FLUX 3 Action also plans further ahead: each call generates 32 actions at 15 Hz, covering 2.13 seconds of motion, compared with 1.0 second for π0.5.

Built on the FLUX 3 backbone

FLUX 3 Action is derived from Black Forest Labs’ broader FLUX 3 multimodal backbone, which was pretrained on image, video and audio data, with video making up more than 95% of training tokens. A second training stage mixed general pretraining data with action-aligned footage, including game recordings, egocentric hand video and teleoperated robots across 14 different robot body types. Black Forest Labs says this pretraining step matters: training on DROID robot data alone kept RoboLab scores below 1%, while adding the broader pretraining raised the same setup to 11.6%.

Outside simulation, Positronic Robotics ran a blind evaluation on a physical Franka arm across 10 DROID-style tasks with three attempts each. FLUX 3 Action completed 28 of 30 attempts, ahead of Cosmos 3 Nano’s 27 of 30, DreamZero’s 20 of 30 and π0.5’s 13 of 30. Black Forest Labs also tested a hybrid setup pairing FLUX 3 Action with the GPT 6 Astra reasoning model, which could execute, edit or override the policy’s predicted actions; the hybrid solved 90% of episodes at lower cost and time than Astra running alone at maximum reasoning effort, which solved 100%.

The DROID policy needs roughly 32GB of GPU memory in BF16 on an H200 chip, or fits on 24GB consumer cards with FP8 quantization and offloading of the text encoder. Black Forest Labs has integrated the model into Hugging Face’s LeRobot framework with Nvidia and published fine-tuning recipes, including one that learned a pick-and-place task from about 200 demonstrations. The model outputs joint targets only, with no built-in velocity, force or workspace limits, leaving safety enforcement to whoever deploys it. Independent testing beyond the Positronic Robotics evaluation has not yet surfaced.