Black Forest Labs Releases FLUX 3 Action, an Open 7B Model for Robots

Black Forest Labs, the studio known for the FLUX image generator, released FLUX 3 Action, an open world action model, on September 22. It has 7B parameters. According to the company's official blog, it outperformed Cosmos 3 Nano, previously the best open model, on the NVIDIA RoboLab-120 leaderboard, even though it is noticeably more compact.

The original Jiqizhixin article ran under a headline about Fei-Fei Li's team and a world model in which swapping the "底座" (base model) sharply raised robot success rates. The available excerpt, however, is devoted entirely to the Black Forest Labs release and does not explain how the two stories are connected. What follows covers only what that excerpt confirms.

What is known about the model

FLUX 3 Action belongs to the class of world action models. These systems combine two tasks: predicting how the surrounding scene will change and generating robot actions. The two tasks are usually handled by separate components. Here they are combined in a single model.

For Black Forest Labs, this is a step outside its usual territory. The company was known as an image generation developer, and now it has moved into physical AI. It has a head start: experience training large vision models bears directly on how a machine "sees" and predicts a scene.

Key details from the available data:

• developer: Black Forest Labs;
• name: FLUX 3 Action;
• type: open world action model;
• size: 7B parameters;
• benchmark: NVIDIA RoboLab-120;
• result: higher than Cosmos 3 Nano, the previous leader among open models;
• fewer parameters than Cosmos 3 Nano (the exact ratio is cut off in the available excerpt).

Why the comparison with Cosmos matters

Cosmos is part of NVIDIA's line of world models, and these models have become the reference point for open-source robotics work. Beating such a model on NVIDIA's own benchmark is a strong move: the comparison takes place on a platform set by the leading hardware supplier for robotics.

Model size matters too. Seven billion parameters fit on a single modern GPU. That makes it easier to fine-tune for a specific robot and to run on board, where memory and power budgets are tight. If independent runs confirm the lead, it would support the argument that for manipulation, the quality of the visual and physical representation of a scene matters more than model size.

The caveats are obvious. The result comes from the company's own blog, and RoboLab-120 measures success in simulation or on a limited set of tasks. How the model will behave on real hardware, with sensor noise, wear and unpredictable objects, remains unclear.

What to check before using it

A team considering FLUX 3 Action should take a few steps:

• read the original blog and the model card: license, weights, input and action formats;
• reproduce the RoboLab-120 result on your own hardware rather than relying on the reported figures;
• check compatibility with your manipulator and control frequency, since world action models are sensitive to latency;
• run tasks outside the benchmark, especially with rare objects and poor lighting;
• compare the fine-tuning cost of a 7B model with that of larger alternatives.

If you are interested specifically in the work by Fei-Fei Li's group mentioned in the headline, consult the primary source: the available text contains no details about it, and inventing them would be wrong.