FLUX 3 Action: Black Forest Labs' Open Model for Robots

Robotics has long run into the same wall. A model can describe a kitchen, a warehouse shelf or a workbench in detail, but that does not mean it knows what a robot arm should do next. Describing a scene is one skill. Acting in it is another.

Black Forest Labs, often shortened to BFL, is going after the second skill with FLUX 3 Action, a new open model built for robots. The company describes it as a world-action model, and its pitch rests less on size than on doing more with fewer parameters and less time.

From watching to acting: what a world-action model does

FLUX 3 Action builds on FLUX 3, a multimodal model from the same company. According to BFL, FLUX 3 was trained mainly on video, with image and audio data added.

That foundation matters. Video carries information that still images cannot: how objects move, what happens after something is pushed, and what order events follow. A model that has seen a great deal of footage has a better chance of learning how the physical world tends to behave.

The robotics version applies this to a practical task. It takes in observations, such as video from several cameras around a robot's workstation, and predicts two things. The first is the action the agent will carry out next. The second is how the surroundings will change because of that action.

In short, the model links a decision to its result. It does not only judge what a scene looks like now. It also estimates what the scene will look like once the robot has acted.

Smaller and faster: the RoboLab-120 results

BFL points to RoboLab-120, a leaderboard for robotics models, to back its claims. The company says FLUX 3 Action set a new best success rate there.

The more notable figure is the model's size. FLUX 3 Action has seven billion parameters. According to BFL, that is less than half the size of the previous best open model. The company also says it runs up to 3.95 times faster.

These numbers come from BFL itself, so independent testing will show how well they hold up. Even so, the direction is clear. The company is betting that efficiency, not scale, is the more useful goal for machines that have to move in the real world.

Why speed matters when the machine has a body

BFL's reasoning is simple. Large reasoning models can be good at planning, the company says, but they are often too slow and too large for robots.

A chatbot can take a few extra seconds to answer without much harm. A robot arm in the middle of a task has less room to wait. If the model is still working out its next move after the scene has already changed, the plan may no longer fit the situation.

Size creates a second problem. BFL notes that efficiency is also key for running models locally. A robot that has to send every decision to a remote data centre depends on that connection. A model small enough to run on the machine, or close to it, avoids that. This is one reason more organisations are looking at bringing open-weight models in-house instead of relying only on hosted services.

Beyond the factory floor: games and computer use

BFL does not limit FLUX 3 Action to physical machines. The company also sees digital environments as a field of use.

Video games already serve as a test ground for navigation. A game world offers many of the same challenges as a physical space: moving through rooms, avoiding obstacles, reacting to change. But mistakes cost nothing, and a test can be repeated as often as needed.

Further out, BFL mentions agents that react quickly and operate computers. That puts the model near a wider push across the industry, where products such as Meta's Muse with Mac control aim to let AI click, type and navigate software for the user. Computer-using agents also raise their own questions, since the more an agent can do, the more carefully developers need to think about where the security boundary sits.

Open weights, open questions

BFL has published the weights of FLUX 3 Action on Hugging Face. Researchers and developers can download the model, test it on their own hardware and build on it.

For robotics, that openness has practical value. Teams can check the company's benchmark claims against their own tasks and adapt the model to setups that BFL never tested.

The larger question is whether a compact model trained mostly on video can bridge the gap between seeing and doing across many kinds of robots and tasks. BFL's leaderboard results are a starting point. The answer will come from what developers build with it now that the weights are public.