FLUX 3 Action predicts robot movements and their visual outcomes together

Black Forest Labs has released FLUX 3 Action, a 7-billion-parameter open-weight World Action Model for robotics. Its variants reach up to 42.2% success on RoboLab-120 in the team's reported evaluations, with faster inference through distillation.

From video generation to robot control

A robot needs to anticipate the consequences of its movements, execute part of a plan, and observe its environment again. FLUX 3 Action applies the multimodal FLUX 3 backbone to this control loop.

Given an instruction, camera observations, and joint states, the model jointly predicts motor commands and the corresponding future video. Both outputs come from the same model rather than a separate video generator and action decoder. Video-heavy pretraining

The 7B model builds on multimodal pretraining with images, video, and audio, with video accounting for more than 95% of training tokens. An additional training stage introduces action-aligned data from gaming, egocentric human recordings, handheld grippers, and teleoperation across 14 robot embodiments.

The model is subsequently fine-tuned for a target robot and its action space. Black Forest Labs reports that multimodal pretraining improves action learning when dedicated robotics data is limited. Faster inference, with a measurable trade-off

On RoboLab-120, the guidance-distilled variant reaches 42.24% success in FP8. The single-step variant reaches 37.92%, compared with 36.8% for Cosmos 3 Nano in the published comparison.

Depending on hardware and precision, Black Forest Labs reports a 1.52× to 3.95× speedup over Cosmos 3 Nano. Single-step inference further reduces computation, at the expense of some task success. These figures concern processing time relative to predicted robot-motion duration, rather than necessarily the latency of each individual call.

In a separate physical-robot evaluation conducted by Positronic Robotics, FLUX 3 Action completed 28 of 30 attempts across ten DROID tasks on a Franka arm. The result provides a real-world measurement, although the test covers a limited number of attempts and tasks. Fast control meets slower reasoning

Black Forest Labs also tested a hybrid system pairing FLUX 3 Action with GPT-6 Astra. The action model proposes movements, while Astra can approve, modify, or replace them when additional planning is required.

In the published simulation protocol, the low-effort Astra hybrid completed 45 of 50 episodes, with an estimated cost of $8.77 and an average time of 8 minutes and 8 seconds per success. Pure Astra at maximum reasoning effort completed all 50 episodes, with higher cost and longer execution time. These estimates are specific to the evaluation setup.

The FLUX 3 Action weights are available as an open-weight release. The accompanying report also explores games as a testing environment for action prediction beyond physical robot manipulation.