Black Forest Labs’ September 22 FLUX 3 Action release puts robot control inside a seven-billion-parameter world-action model. The open-weight system predicts actions together with the video it expects those actions to produce. Its report asks whether a model can preserve useful physical-task performance while reducing the processing cost of control. BFL’s report
The model combines an instruction, camera observations and robot state to predict motor actions and their visual outcomes. This connects prediction to action: the system anticipates what the environment should look like after its proposed movement. The report presents this as a route from multimodal pretraining to robotics, with digital environments as another research direction.
BFL reports 38.3% success on RoboLab for its single-step checkpoint and 42.2% for a guidance-distilled checkpoint, compared with 36.8% for Cosmos 3 Nano in the cited comparison. Those are reported benchmark results, not a demonstration that the model can reliably complete arbitrary physical work. The different checkpoints also show that speed and task success remain separate choices.
One speed comparison needs particular care. BFL reports a 1.34-to-2.28-times improvement over Pi0.5 in processing time per second of robot motion. The source explicitly distinguishes that measure from latency per call: FLUX 3 Action predicts a longer movement horizon. Readers comparing interactive responsiveness should not silently substitute one measurement for the other.
The report also describes a physical evaluation conducted by Positronic Robotics. According to BFL, a Franka arm attempted ten DROID tasks three times each, with a blinded operator and equal 240-second windows. FLUX 3 Action completed 28 of 30 attempts. This offers a concrete physical test alongside simulation, although the small task set and BFL’s account of it do not establish broad industrial reliability.
My inference is that a useful action model must be judged on what it successfully completes, how quickly it can reconsider, and what happens when its prediction is wrong. Lower processing cost can expand the room for repeated observation or higher-level planning. The release supplies evidence for that research tradeoff; transferring the result to an unfamiliar robot, environment or task remains a separate question.