Stay connected with KayaToday, follow us on Instagram and Facebook for the latest news and reviews delivered straight to you.
The hardest problem in robotics AI is not making a model that performs well in a lab. It is making one that developers can actually take home, adapt to their own hardware, and trust to work reliably in a factory or warehouse. Black Forest Labs, the German startup that built its reputation on image generation, is now making a direct bid at that problem with FLUX 3 Action, a 7-billion-parameter open-weight model that the company says can translate camera footage, robot state data, and plain-language instructions into physical actions.
The headline claim is a 42.92% overall success rate on NVIDIA’s RoboLab-120 benchmark, which BFL says places it above every model currently listed on that public leaderboard, including NVIDIA’s own Cosmos3-Nano-Policy, which scored roughly 6.1 percentage points lower while running on 16 billion parameters. That size gap matters: FLUX 3 Action uses just 44% as many parameters as its closest benchmark rival and runs 1.43 times faster by BFL’s own measurements. The company plans to release the weights, code, fine-tuning recipe, and reproducible examples, including an implementation built around the relatively affordable SO-101 robot and Hugging Face’s LeRobot framework.
What the Benchmark Actually Proves, and What It Does Not
RoboLab-120 is an NVIDIA-designed simulation benchmark covering 120 tasks that test visual understanding, relational reasoning, and procedural skills across multiple difficulty levels. NVIDIA built it partly to avoid the saturation problem that has made older robotics benchmarks less informative as models have improved. A 42.92% score on a genuinely difficult benchmark is not trivial, and beating both Cosmos3-Nano-Policy and the previous overall leader OASIS WAM, which sat at 39.0% before this release, is a meaningful result.
But the competitive landscape is more complicated than a single leaderboard position suggests. Ai2’s MolmoAct 2 runs on roughly 5 billion parameters, and NVIDIA distributes a 3-billion-parameter GR00T N1.7 checkpoint. Neither appears alongside FLUX 3 Action on RoboLab because they are evaluated on different mixes of real-world and simulated tasks. Ai2 reports MolmoAct 2 achieving an 87.1% average success rate across five of its own real-world Franka evaluation tasks, a figure that cannot be compared directly with RoboLab percentages because the task sets and methodology differ entirely. This is a recurring frustration in robotics research: hardware, task distributions, and evaluation environments vary so dramatically between groups that no single number cleanly establishes an overall leader.
BFL also emphasises inference speed, using real-time factor as the key metric, which measures how much compute time a model spends deciding what to do relative to the amount of robot motion it controls. The company says distilled FLUX 3 Action variants establish a new Pareto frontier between RoboLab success rate and inference speed across datacenter GPUs, and that the model surpasses Physical Intelligence’s π0.5 in real-time factor while maintaining higher task success. These are genuinely important deployment metrics because a policy that cannot produce actions fast enough will cause pauses or jitter in physical operation. The caveat is that BFL supplied these results ahead of publication and NVIDIA’s public leaderboard had not yet been updated to reflect them when checked before launch.
The Architecture Behind the Claims
FLUX 3 Action is not a conventional vision-language-action model. BFL and NVIDIA classify it as a World Action Model, or WAM, meaning it links predictions about how the world will evolve with action generation rather than treating the two as separate outputs. The architecture builds on the image, video, and audio pretraining behind the broader FLUX 3 family, but uses a smaller configuration optimised for practical deployment.
During training, the model learns to predict future video frames and actions simultaneously. At inference time it takes recent camera frames, the current system state, and a task description, then returns the next 32 actions alongside a prediction of how the visual scene will change. The robot observes the updated environment and repeats the loop. This approach traces back to BFL’s Self-Flow research, which attempts to make a generative model learn useful representations of its inputs at the same time it learns to generate outputs, eliminating dependence on separate frozen representation models such as CLIP or DINO.
The practical implication is that BFL is betting a model pretrained deeply on video will absorb useful representations of motion, contact, object behaviour, and cause-and-effect, and that those representations can then be adapted for robotic control without enormous amounts of robot-specific demonstration data. BFL says the policy shown in its new demonstrations was fine-tuned from around 200 teleoperated episodes covering a handful of related pick-and-place tasks, and that the objects shown in test videos were not present in that training set. One supplied demo shows the robot making an initial mistake and then attempting the task again, behaviour the researchers described as “the model failed first, and then it tried again and better.” BFL is careful to characterise this as a selected demonstration rather than evidence that autonomous recovery will generalise broadly, which is the right level of caution.
Open Weights in a Crowded Field, and What Comes Next
The open-weight release is arguably more consequential than the benchmark position. The robotics AI space now has a clear spectrum. At the proprietary end, Google’s Gemini Robotics On-Device 2, released in July, is designed to run locally on robot hardware but is distributed only to selected testers. In the middle, NVIDIA’s GR00T N1.7, released in April, is open and can be post-trained for specific robots and tasks. At the most open end, Ai2’s MolmoAct 2 has released not just weights but training code, fine-tuning scripts, datasets, evaluation rollouts, and LeRobot integration.
FLUX 3 Action sits closer to the Ai2 end of that spectrum, though BFL has not yet published final commercial license terms for the weights, which matters for any team considering production deployment. There is also no FLUX 3 Action API announced yet. BFL’s existing public product page still describes Action as rolling out through selected research and commercial partners, making self-hosting and fine-tuning the central proposition for now.
For developers in Malaysia, Singapore, and across Southeast Asia building robotics applications in manufacturing, logistics, or agriculture, the practical appeal of an open-weight model is straightforward. Teams can run inference within their own infrastructure, inspect and modify the model, and keep proprietary robot demonstrations and operational data inside their own environment rather than routing it through a hosted service. The 7-billion-parameter size also means the compute requirements are more accessible than a 16-billion-parameter alternative, which reduces the barrier for smaller teams and research groups.
BFL’s trajectory since emerging in 2024 has been unusually rapid. The company raised a $300 million Series B at a $3.25 billion post-money valuation in December 2025, bringing total funding to more than $450 million, and has expanded from image generation through video and audio into robotics within roughly 18 months. The benchmark numbers suggest the world-model approach can be competitive in simulation. The real test is whether outside teams, working with their own robots and their own data, can reproduce those gains in production. That is the question FLUX 3 Action’s open release is now positioned to answer.
Read More: Your AI Agent Is Quietly Forgetting Its Own Rules, and a Bigger Context Window Will Not Save You
