Black Forest Labs has released FLUX 3 Action, an open-weights 7B parameter World Action Model (WAM) designed for robot control. Unlike traditional Vision-Language-Action (VLA) models, a WAM predicts both future actions and visual outcomes jointly, allowing for simulation of environmental changes before physical execution to reduce trial-and-error costs.
Model ReleasesPartnershipsBlack Forest LabsFLUX 3 Action
Black Forest Labs Releases FLUX 3 Action, a 7B World Action Model for Robot Control
In the RoboLab-120 benchmark, FLUX 3 Action achieved a success rate of 42.92%, surpassing the previous leaders: NVIDIA's Cosmos 3 Nano (36.8%) and the open-source pi0.5 (28.0%). Notably, FLUX 3 Action achieves these results with fewer parameters and operates up to 3.95x faster than Cosmos 3 Nano on various hardware configurations.
The model benefits from a multimodal training approach using the "Self-Flow" method, which incorporates image, video, and audio data to improve generalization. To improve deployment efficiency, Black Forest Labs utilized knowledge distillation to create a single-step version of the model.
Black Forest Labs has collaborated with NVIDIA to integrate FLUX 3 Action into the open-source robotics platform LeRobot, providing recipes for fine-tuning and support for deployment on NVIDIA Jetson edge devices. The company noted that while pure action models are efficient, combining them with reasoning models like GPT 6 Astra can further enhance performance for complex, long-horizon tasks.
Sources
- ロボット制御向けAIモデル「FLUX 3 Action」をBlack Forest Labsが公開 （GIGAZINE, 2026-09-25）
- Black Forest Labs