FLUX 3 Action: Open 7B World-Action Model Tops RoboLab
- Source
- x.com
- Date
Black Forest Labs, known for its FLUX image models, has moved into robotics with FLUX 3 Action, a 7-billion-parameter world-action model that jointly predicts video and actions. The company says it tops the RoboLab benchmark while using 56% fewer parameters than the previous best open vision-language-action model and running up to 3.95 times faster. The release is open: weights, fine-tuning recipes, and integration with NVIDIA tooling and Hugging Face LeRobot ship together. That combination matters more than the benchmark rank. Robotics teams usually face a choice between large closed policies and smaller open models that underperform. A smaller open model that leads on a public benchmark and comes with a fine-tuning path lowers the cost of adapting a policy to a specific robot, and the parameter savings make edge deployment more realistic. The benchmark result is the company's own claim, so independent reproduction on RoboLab is the next thing to watch.
Introducing FLUX 3 Action. An open weights 7B World Action Model that achieves first place on the RoboLab benchmark. It outperforms the previous best open model by 6.1 percentage points while using 56% fewer parameters and running up to 3.95x faster. FLUX 3 Action removes the usual trade-off between world action model performance and VLA speed: it still predicts video and actions together, but plans more than twice as far ahead and runs faster per second of robot motion than the strongest open VLA. Teams can fine-tune FLUX 3 Action on their own demonstrations to create policies for a particular robot and task. Together with @nvidia, we also integrated FLUX 3 Action natively into @huggingface's LeRobot, with fine-tuning recipes included and edge deployment on NVIDIA Jetson. Beyond robotics, we’re also seeing promising results training task-specific policies for acting in simulated environments like gaming, controlling a vehicle, computer use, and wherever else a model needs to understand a visual environment and then choose what to do next. FLUX 3 Action builds on the same image, video, and audio pretraining as FLUX 3, but uses a smaller architecture designed for practical…
Context
Black Forest Labs' FLUX 3 Action is an , 7-billion-parameter robot-control model built to close the gap between two kinds of policy. World action models (WAMs) predict future video frames together with the robot's actions, which lets them carry over what they learned from large-scale video. Vision-language-action models (VLAs) are the faster alternative. The company's technical report says WAMs lead the RoboLab simulation , with Nvidia's Cosmos 3 Nano succeeding on 36.8% of tasks against 28.0% for Pi0.5, the strongest open VLA, but that Cosmos 3 Nano needs about 4.7 times as much processing per second of robot motion.
The announcement claims FLUX 3 Action keeps the WAM approach while removing most of that cost: first place on RoboLab, 6.1 percentage points above the previous best open model, with 56% fewer parameters and up to 3.95x faster. The report fills in the comparison. The previous best open model is the 16-billion-parameter Cosmos 3 Nano, and the 3.95x figure is the top of a 1.52x to 3.95x range across consumer, workstation and datacenter GPUs. Its leaderboard table lists FLUX 3 Action at 42.92%, which matches the 6.1-point margin, while the report's text gives 42.2% for its best checkpoint and 38.3% for a faster single-step checkpoint.
The speed comparison with Pi0.5 is measured per second of robot motion, not per call. Each prediction covers 2.13 seconds of motion against 1.0 second for Pi0.5, which is what the post means by planning more than twice as far ahead, and the report puts the resulting advantage for its faster single-step checkpoint at 1.34x to 2.28x on workstation and datacenter GPUs.
Black Forest Labs says it is releasing the weights, code and recipe so teams can adapt the model to their own robot and task, and that with Nvidia it integrated the model into Hugging Face's LeRobot, with deployment on Nvidia Jetson hardware. The benchmark figures are the company's own. The post describes uses in gaming, vehicle control and computer use only as promising early results. The weights page and the attached video were not reviewed.
- open weights — A model whose trained parameters are published for anyone to download and run — unlike API-only models you can access but never possess.
- benchmark — A standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.
- fine-tuning — Taking a trained model and training it a bit more on your own examples so it gets better at one specific job.
Checking sign-in…
Loading comments…