Black-forest-labs released FLUX 3 Action on September 23, 2026, saying it predicts future frames and actions together. It is the company's first world action model update since the release of FLUX.2.
Black-forest-labs reported 42.92% success on the RoboLab benchmark, measured on the RoboLab-120 dataset. That compares with 36.8% for the 16B Cosmos 3 Nano policy.
FLUX 3 Action is built on a diffusion transformer and targets robotic control and game environments. Availability begins with open-source release, initially for researchers and developers.
"FLUX 3 Action is an open weights world action model (WAM); it predicts future frames and actions together," said Patrick Esser, Black-forest-labs engineer. The model is trained on action data from several robot embodiments and uses a frozen video VAE to encode frames.
The announcement follows collaboration with NVIDIA on PEFT fine-tuning recipes and edge deployment. Black-forest-labs said the model's integration with LeRobot enables policy training for specific tasks.
Black-forest-labs did not say how the model performs in unseen environments, and raised the open question of whether the model can adapt to new camera positions. The company said the model recovers from its own mistakes and works when camera positions change.
Source: huggingface
