World Action Model Cheat Sheet: FLUX 3 Action

Chatbot vs world action model

What it is

• An open weights, 7B parameter world action model from Black Forest Labs, released September 23, 2026.
• It "predicts future frames and actions together." Actions are a second token stream in the same sequence as the video.
• The state vector is the robot's joint positions, or the last action taken in a game.

Key numbers

• 42.92% success on the RoboLab benchmark, first place. The 16B Cosmos 3 Nano policy scores 36.8%.
• 7B vs 16B parameters: less than half the size of Cosmos 3 Nano.
• ~200 teleoperated episodes were enough to adapt it to pick-and-place on an SO-101 arm.

Beyond robot arms

• Games: Black Forest Labs trained policies for two custom games on 800 episodes per game. It beat the scripted bot on kills in GRUNT (15 vs 13) and came close in VECTOR (3.3 vs 3.4 km).
• Drone: it was trained on 800 simulated flights in NVIDIA Isaac Sim. It outputs forward, lateral, up and yaw.

How to try it

1. Get the weights from the FLUX 3 Action collection on Hugging Face (start with black-forest-labs/flux-3-action-base).
2. Get the code at github.com/black-forest-labs/flux-action.
3. Use the LeRobot integration. It turns FLUX 3 Action into a policy class that you train and run with LeRobot's own tools.
4. Record your own demonstrations and fine-tune on them. Black Forest Labs says "the fine-tuning recipe is just as important as the weights." The config, dataset module and pitfalls are in the BFL docs.
5. For edge deployment, Black Forest Labs lists NVIDIA Jetson support.

Before you ship

The weights are released under the FLUX Kommunity License v1.0. Read its terms before any commercial use.

Sources: huggingface.co/blog/black-forest-labs/flux-3-action · bfl.ai/models/flux-3-action · VentureBeat