Black Forest Labs Details Robot Model, Claims Top Score Ahead of Public Release

A robot model once limited to early-access partners now comes with a performance claim. Its maker, Black Forest Labs, says the 7-billion-parameter FLUX 3 Action model scored 42.92% on a robotics simulation benchmark, topping the previous public leader. The result gives developers a reason to look closely, but the promised model weights are not yet public.

From camera view to robot movement

FLUX 3 Action takes recent camera images, the robot’s current state and a task description. It then produces the next 32 actions alongside a prediction of how the scene will change. After moving, the robot observes its surroundings again and repeats the process. The aim is to make each set of movements responsive to what the camera sees next, rather than treating the entire task as a fixed sequence.

The design draws on Black Forest Labs’ work with images, video and audio. During additional training, the Action model learns to predict future video frames and robot actions together. It is an attempt to put knowledge of how a scene changes to work controlling a machine. The company calls this a “World Action Model”; the useful distinction is that it predicts both what might happen next and what the robot should do.

The simulation result, in context

Black Forest Labs says FLUX 3 Action achieved a 42.92% overall success rate on NVIDIA’s RoboLab-120, a simulated test spanning 120 robotics tasks. That would beat the 39.0% score of OASIS WAM, which held the top overall position immediately before this result. Black Forest Labs also puts its model 6.1 percentage points ahead of NVIDIA’s Cosmos 3. Those are company-supplied results: NVIDIA’s public leaderboard had not yet added FLUX 3 Action when VentureBeat checked before publication.

42.92%FLUX 3 Action success rate

Black Forest Labs reports this overall success rate on NVIDIA’s simulated RoboLab-120 tasks; the public leaderboard had not yet been updated with its result.

The size and speed comparison is narrower than a claim to be the smallest robot model. Black Forest Labs says Action has 7 billion parameters, against 16 billion for NVIDIA’s Cosmos3-Nano-Policy, and runs 1.43 times faster. Other robotics models are smaller still, but they are tested on different mixes of simulated and real-world work. A RoboLab lead therefore does not establish which model will perform best across robots, hardware and tasks.

What the demonstrations can show

For its robot demonstrations, Black Forest Labs says it fine-tuned a policy using about 200 episodes in which a person remotely controlled a robot through related pick-and-place tasks. The company says the objects in the test videos were absent from that training set. In one company-selected clip, the robot fails an attempt and tries again. That is a useful glimpse of the behavior the company wants, not proof that recovery will work reliably in other settings.

The company has also tested the approach beyond robot arms. It says task-specific versions flew real drones and finished the computer game Doom without a death. Black Forest Labs describes these as early experiments, not production capabilities. Neither example settles whether a team can adapt the model to an unfamiliar robot and keep it performing consistently outside a selected demonstration.

The release developers are waiting for

This is not Action’s first appearance. In its July FLUX 3 announcement, Black Forest Labs said the Action version was available for early access and that broader availability and open-weight versions would follow. The new account adds specific benchmark and design details, but it should not be mistaken for a public download. Black Forest Labs told VentureBeat it plans to release weights, code, a fine-tuning recipe, benchmarks and reproducible examples, including one built around the SO-101 robot and Hugging Face’s LeRobot framework.

The public weights and their license terms remain unavailable, and Black Forest Labs has not announced an Action API or its pricing. Those gaps matter to teams considering whether they can fine-tune and run the model themselves, especially for commercial use. For now, the benchmark describes a promising simulation result; the planned release would let outside developers test how much of that promise survives on their own machines.