In brief

Black Forest Labs released FLUX 3 Action, a 7-billion-parameter model that turns camera images and text instructions into robot movements. It leads NVIDIA’s RoboLab-120 leaderboard with 42.9%. The weights are free for non-commercial use, and most speed and real-robot figures come from the company.

New to this? Read it in simple words
  • Black Forest Labs, a company known for image AI, released a model that controls robots.
  • It came first on a public robot test that runs in a computer simulation.
  • It succeeded in 42.9% of attempts, so it still fails more often than not.
  • It is free for non-commercial use. Most other claims come from the company.
Words to know
Simulation
A computer-made world where robots can be tested safely before real use.
Open weights
The trained model files are published, so others can download and run them.
Leaderboard
A public ranking that compares how well models do on the same test.

What the model does

FLUX 3 Action has 7 billion parameters. It takes camera images, the robot’s current state, and a written instruction. It then predicts the robot’s next movements, together with the next video frames.

The company calls this a world action model: it imagines what the camera will see next while it plans the next moves.

The weights are on Hugging Face under the FLUX Kommunity License. They are free for non-commercial use. Most commercial use needs a separate licence from the company, which may cost money.

Sources124

LEADERBOARD CHECK 01
First place in simulation, at less than half the size.

Success rates on the 120 simulated tasks of RoboLab-120. Real-robot results come from the company.

First place in a public test

RoboLab-120 is a set of 120 simulated robot tasks, such as picking up, stacking and using tools. Its leaderboard is hosted by NVIDIA Research.

On that leaderboard, FLUX 3 Action ranks first, with 515 successes in 1,200 attempts, or 42.9%. Next come HiDream-O1-Embodied at 39.9% and Atomic-WAM at 39.6%. NVIDIA’s own Cosmos3-Nano-Policy scores 36.8%.

The company says its model is less than half the size of Cosmos3-Nano-Policy, which has 16 billion parameters.

The leaderboard does not show when entries were added. We could not confirm whether NVIDIA ran this result again itself.

Sources345

Claims still to check

Black Forest Labs says its model runs up to 3.95 times faster than Cosmos 3 Nano. Reports have given different speed figures, so we treat all speed numbers as company claims.

The company also reports a test on a real robot arm by Positronic Robotics. Its model succeeded in 28 of 30 attempts, against 27 for Cosmos 3 Nano. We found no report from Positronic itself.

MarkTechPost reports that the model can run on a 24-gigabyte consumer graphics card when compressed. That would put it within reach of small labs.

Simulation scores are only a first step. A 42.9% success rate means the model still fails more tasks than it completes.

Sources145

Sources

Every fact in this story comes from the sources below. Open them to check our work.

  1. 1
    Primary source · September 2026FLUX 3 Action: A 7B World Action Model for Robot Control Black Forest Labs
  2. 2
    Primary source · September 2026black-forest-labs/flux-3-action-base Hugging Face
  3. 3
    Research · September 2026RoboLab leaderboard NVIDIA Research
  4. 4
  5. 5
How we checked this story

We read the company’s model page, its Hugging Face card and licence, and checked the RoboLab leaderboard ourselves. We compared MarkTechPost and VentureBeat. Speed figures differ between reports, so we give the company’s own numbers and label them as claims.