FLUX 3 Action takes first place on RoboLab at 7B
Black Forest Labs put numbers on its open weights world action model, and Animesh Garg pulled out the ones that matter.
Black Forest Labs says FLUX 3 Action takes first place on the RoboLab benchmark, beating the previous best open model by 6.1 percentage points while using 56% fewer parameters and running up to 3.95x faster. The model is a 7B world action model, meaning it predicts future video frames and robot actions together rather than picking one. Weights, code, the fine-tuning recipe, benchmarks and reproducible examples are all out, which is the part that lets anyone check the claim.
The pitch is that you usually pay for that video prediction in latency. Black Forest Labs says FLUX 3 Action plans more than twice as far ahead as the strongest open VLA (a model that turns camera images and text into robot actions) and still runs faster per second of robot motion. It builds on the same image, video and audio pretraining as FLUX 3, with a smaller architecture, and in midtraining it learned to predict actions and future frames at the same time.
Garg fills in the comparison
Animesh Garg named the model it is measured against. He says FLUX 3 Action beats Cosmos 3 Nano, a 16B model, on RoboLab-120 by 42.9% to 36.8%, at less than half the parameter count and up to 3.95x faster. He also points at the part the announcement skips, that pairing the model with a high level reasoner cuts execution costs by 29% and runtimes by 40% compared with pure reasoning, letting the fast policy handle motor primitives and calling the LLM only to correct course.
The release goes further than a checkpoint on a hub. Black Forest Labs says it worked with NVIDIA to integrate the model natively into Hugging Face's LeRobot, with fine-tuning recipes and edge deployment on NVIDIA Jetson. Teams fine-tune on their own demonstrations to get a policy for their robot and task. The company also says it is seeing promising results on non robot acting, including games, driving a vehicle and computer use.
Shown versus claimed. Every number here comes from the company's own benchmark run, plus Garg's reading of it. Nobody outside has evaluated it yet, and the earlier release post did not come with independent scores either. The open weights at least make that check possible.
Everyone is pivoting into a robotics neolab
pairing F3A with high-level reasoners cuts execution costs by 29% and runtimes by 40% over pure reasoning
Sound like you were at the demo
One email a day. Under three minutes. The demos, the deployments and the part the press release left out.