Updated

FLUX 3 Action

Black Forest Labs released FLUX 3 Action, an open-weight 7B world action model that predicts future actions from camera frames and text instructions. Black Forest Labs reports that a DROID fine-tune placed first on the RoboLab-120 benchmark at a 42.92% success rate.

By No Code MBA · Facts checked

Provider
Black Forest Labs
Announcement date
Status at last check
Generally available
Model family
FLUX 3
Capabilities
Image, Video

What changed

FLUX 3 Action is a 7B diffusion transformer world action model that takes camera frames, a robot state vector and a text instruction, and outputs 32 future actions plus optional predicted video frames. It uses a video VAE and Qwen3-VL-4B as the frozen instruction encoder, and was pretrained on multi-embodiment robot data. Black Forest Labs reports a DROID fine-tune ranks first on RoboLab-120 at 42.92% success, ahead of the 16B Cosmos3-Nano-Policy at 36.8%.

Access and availability

Base weights are on Hugging Face at black-forest-labs/flux-3-action-base (gated, requires accepting terms) under the FLUX Kommunity License v1.0, with DROID and SO-101 checkpoints via LeRobot and code at github.com/black-forest-labs/flux-action.

Official access documentation

What you can build

Use FLUX 3 Action through the LeRobot integration to fine-tune a robotics policy for your own robot arm or test it in simulation through NVIDIA Isaac Sim before deploying on hardware.

Limitations

Black Forest Labs says the base is an adaptation component rather than a complete robot policy, so new embodiments need their own action heads and fine-tuning. Nothing in the model bounds joint velocity, force or workspace, so applications must enforce safety limits. The license's commercial terms are unclear, and Black Forest Labs has published no latency figures.

Sources and release history