Black Forest Labs launches FLUX 3 Action, an open robotics AI model
Black Forest Labs (BFL) has released FLUX 3 Action, an open AI model built specifically for robotics. It is built on top of the company's multimodal FLUX 3 model, which was trained mainly on video but also on image and audio data. FLUX 3 Action works as a world-action model: it takes in multi-camera video feeds of a robot's workspace, then predicts what action the robot should take next and how the environment will change as a result.
According to BFL, FLUX 3 Action sets a record for success rate on the RoboLab-120 leaderboard while using just seven billion parameters, less than half the size of the previous best open model on that leaderboard, and running up to 3.95 times faster than it. BFL argues that large reasoning models can plan well but tend to be too slow and too bulky to run on robots, which is why this combination of a small parameter count and high speed matters for deployment directly on robotic hardware.
Beyond physical robots, BFL says it sees potential for the same approach in digital environments: video games could serve as testing grounds for navigation, and the company suggests that fast-reacting agents able to operate computers could follow down the line. The model's weights are available on Hugging Face.
Key facts
- FLUX 3 Action is an open robotics model from Black Forest Labs, built on the multimodal FLUX 3 model trained mainly on video plus image and audio data.
- It takes multi-camera video of a robot workspace and predicts the next action and the resulting change to the environment.
- BFL says it sets a record for success rate on the RoboLab-120 leaderboard with seven billion parameters, less than half the size of the previous best open model, while running up to 3.95 times faster.
- BFL argues large reasoning models plan well but are too slow and bulky for robots, making this efficiency important for on-device use.
- The model's weights are published on Hugging Face; BFL also floats future use in video games and computer-operating agents.
Why it matters
Robotics has largely relied on large, slow reasoning models that plan well but struggle to run directly on hardware with limited compute and tight latency budgets. FLUX 3 Action is positioned as a smaller, faster alternative built for that specific constraint: BFL says it beats the previous best open model on the RoboLab-120 leaderboard while using less than half its parameters and running up to 3.95 times faster, which if borne out would be a meaningful efficiency gain for a field where model size directly limits what can be deployed on a physical robot.
Who it affects
The release is aimed at robotics researchers and companies building on-device robot control systems, since the model is meant to run directly on robotic hardware rather than in the cloud. It also touches developers interested in agents for digital environments, as BFL points to video games and computer-operating agents as possible future applications of the same world-action approach.
How to use it
FLUX 3 Action's weights are available on Hugging Face, making the model open for researchers and developers to download and experiment with. No pricing, licensing terms, or release date beyond the announcement are given in the source.
How solid is it
The efficiency and success-rate claims, including the seven-billion-parameter count, the size comparison, and the 3.95x speed figure, are attributed to BFL itself rather than to an independent benchmark report, and the identity of the 'previous best open model' it is compared against is not named in the source.
Risks and caveats
Because the comparison figures come from BFL's own account of its RoboLab-120 results rather than a named third-party model or independent verification, the actual margin over competing approaches is hard to assess from the source alone. The source also gives no release date, pricing, or licensing details beyond the availability of weights on Hugging Face.