FLUX 3 Action

A 7B open-weight world action model from Germany's Black Forest Labs, released September 2026, predicting frames and actions together.

FLUX 3 Action is a robot model released in September 2026 by Black Forest Labs, the German company known for its FLUX line of image-generation models, with 7 billion parameters and weights open on Hugging Face. It is a world action model (WAM, a model that jointly predicts future video frames and robot actions), adapted from the company's multimodal FLUX 3 backbone, which was pretrained on large amounts of video, image, and audio data, mostly video. The model takes in camera images, joint state, and a text instruction, and outputs both a robot action and a predicted future frame at the same time, aiming to carry the physical common sense learned from video pretraining over into control. The company says it supports multiple action spaces, including end-effector pose and joint angles, and demonstrated it on a Franka arm and the SO-101. It is an example of an image- and video-generation company moving into embodied AI.

ExampleThe company reports that a guidance-distilled version reaches a 42.2% success rate on the RoboLab-120 simulation leaderboard, ahead of NVIDIA Cosmos 3 Nano's 36.8%, and a 93.3% success rate across 10 real-robot DROID manipulation tasks.

Also called

Black Forest Labs open-weight world action model

Related

World Action Model · Open-weight Model · Cosmos 3 · Fast-WAM · DROID (Distributed Robot Interaction Dataset) · SO-100 / SO-101 Arm

Sources

FLUX 3 Action 模型页（Black Forest Labs） (Chinese)
Black Forest Labs 官网 (Chinese)

As of

2026-09