Published: 24 September 2026
Black Forest Labs published FLUX 3 Action on 23 September 2026, a 7 billion parameter open-weights model for robot control. It scores 42.2 percent on NVIDIA’s RoboLab-120 benchmark, ahead of Cosmos 3 Nano at 36.8 percent while using less than half the parameters. The weights are on Hugging Face under a new licence Black Forest Labs calls the FLUX Kommunity License.
What does FLUX 3 Action do?
It turns a camera feed into robot movement. You give it the images, the robot’s current position and a written instruction, and it returns the next block of 32 movements, roughly two seconds of motion at 15 commands per second. It also predicts what the camera will see next, which is where the video half of the model earns its place.
This is a departure for a company known for image generation. Black Forest Labs calls it a World Action Model, and it is built on the same FLUX 3 backbone as its picture models. The claim behind the design is that a model which has learned to predict video frames already understands how objects move, so it needs less robot data to learn a task. Fine-tuning takes roughly 200 recorded demonstrations.
Three versions shipped together:
- flux-3-action-base, the general checkpoint
- flux-3-action-so101, for the low-cost SO-101 arm via Hugging Face LeRobot
- flux-3-action-droid, trained on the DROID robot dataset
- Licence: FLUX Kommunity License v1.0, with the Qwen3-VL-4B text encoder under Apache-2.0
- Speed: 41.06 milliseconds per block of actions on a B200 GPU
FLUX 3 Action benchmarks against Cosmos 3 Nano
Take this from the table: the gap over the nearest open model is about six points, and it comes with half the parameters and roughly three times the speed.
| What is measured | FLUX 3 Action | Comparison |
|---|---|---|
| RoboLab-120 success rate (simulated robot tasks) | 42.2% | 36.8% for Cosmos 3 Nano |
| Same benchmark, weaker checkpoint | 38.3% single-step | 28.0% for π0.5, 25.7% for DreamZero |
| Parameters | 7 billion | 16 billion for Cosmos 3 Nano |
| Inference speed | 2.85x to 3.15x faster than Cosmos 3 Nano | 41.06 ms per action block on a B200 |
| Real robot test | 93.3%, or 28 of 30 tasks | one Franka arm, tasks not itemised |
| Licence | FLUX Kommunity License v1.0 | Cosmos 3 is NVIDIA open-weights |
These are Black Forest Labs’ own figures, published on the model page, with a reported margin of plus or minus 0.36 points on the headline score. RoboLab-120 is NVIDIA’s simulation benchmark, so the comparison to Cosmos 3 Nano runs on NVIDIA’s home ground. The real-robot result is a separate, much smaller test: 30 tasks on a single Franka arm, with no breakdown of which two failed.
What Black Forest Labs is not saying
The licence is new and the terms are not spelled out. Black Forest Labs describes both non-commercial and commercial options but has not published what commercial use costs or requires. For a robotics company deciding whether to build a product on this, that is the only question that matters, and it has no answer yet.
There is also no comparison against closed models. Every rival in the table is open-weights. Physical Intelligence and Google both have systems in this space that are not benchmarked here, so the claim is that FLUX 3 Action leads the open field, not the field. And the model outputs raw joint targets with no limits on speed, force or reach built in. Black Forest Labs says so plainly and tells you to add your own safety layer, which is the right disclosure and also a warning about how much work is left.
What this means
Worth testing now if you already have a robot arm on a bench. A research group or a small automation team with an SO-101 or a Franka can pull the matching checkpoint and fine-tune it on about 200 demonstrations, which is a weekend rather than a project. At 7 billion parameters it runs on hardware you probably have, and the speed figure means it can actually close a control loop instead of thinking too slowly to be useful.
Safe to ignore for now if you are anywhere else. Without published commercial licence terms you cannot build a product on this, and without a comparison to the closed systems you cannot tell whether the state of the art is here or somewhere else. The wider pattern is worth noting though: an image generation company has walked into robotics and taken the top open slot on the first try, using a video model as the foundation. That is the third release in two months arguing that video prediction is the cheap route to robot control, and the argument is starting to hold.
For more information, visit the official announcement of FLUX 3 Action on the Black Forest Labs site.
