Black Forest Labs has introduced FLUX 3 Action, a 7B parameter open-weights world action model designed for robotics. The model utilizes a multimodal backbone trained on image, video, and audio data, with video accounting for over 95% of training tokens. It performs joint video and action prediction, using distillation to improve processing speeds compared to existing models.
In testing on the RoboLab-120 benchmark, the model achieved a 42.92% success rate, outperforming NVIDIA’s Cosmos 3 Nano by 6.1 percentage points while using 56% fewer parameters. Real-world evaluations on a Franka arm saw the model complete 28 out of 30 tasks, surpassing other models like DreamZero and π0.5.
The model is compatible with NVIDIA Jetson for edge deployment and is integrated into the Hugging Face LeRobot library. It requires 32 GB of GPU memory in BF16 on an H200, though it can run on 24 GB cards using FP8 quantization. The FLUX Kommunity License restricts the model to non-commercial use.
Performance tests indicate that FLUX 3 Action runs 1.52x to 3.95x faster than Cosmos 3 Nano across various GPU configurations. When used in a hybrid control setup with GPT 6 Astra, the model solved 90% of episodes at a cost of $8.77 per success, compared to 100% success for pure Astra at $13.47.
