BFL reports that FLUX 3 Action is a 7B open World Action Model and that it jointly predicts video and actions. In its RoboLab-120 table, BFL lists 42.92% overall success for FLUX 3 Action, versus 36.8% for the 16B Cosmos3-Nano-Policy and 28.0% for π0.5. The company also says its step-distilled variant reduces sampling to a single step without separating video prediction from action prediction.
The mechanism is multimodal pretraining followed by joint video-action training and embodiment-specific fine-tuning, then distillation to remove a guidance pass and reduce sampling steps. BFL argues this retains representations learned from broad video data while lowering the sequence and inference burden of a world-action model. Why now: robot-control stacks are increasingly forced to choose between fast action policies and slower visual world models, making a reported shift in that frontier operationally relevant.
The implication is that a hybrid stack can assign frequent low-latency control to a compact world-action policy and reserve expensive reasoning for exceptions. BFL reports that, in its cited experimental setting, combining FLUX 3 Action with a frontier reasoner reached 90% episode success while lowering cost and time per success relative to pure reasoning. The evidence boundary is that these are vendor-reported results, largely framed around RoboLab simulation and a particular serving setup; they do not establish reliability, safety, or transfer performance across real robots and unreported task distributions.
Watch whether independent real-world evaluations reproduce the claimed speed–success trade-off after embodiment-specific fine-tuning, and whether the released weights and code enable comparable deployments outside BFL’s benchmark configuration.