Black Forest Labs releases FLUX 3 Action, an open-weights robot control model

Black Forest Labs released FLUX 3 Action on September 24, 2026, an open-weights model that turns camera frames, a robot's current state, and a text instruction into a sequence of physical actions. It is the image-generation lab's first move into robotics, adapting the visual understanding behind its FLUX image models to real-world and simulated control tasks.

What's new

The model card states plainly what the system does: "FLUX 3 Action is an open weights 7B world action model. It takes camera frames, the robot's state and a text instruction, and returns the next chunk of actions, denoised together with the next video frames." Rather than outputting a single next move, it predicts a short chunk of upcoming actions together with the video frames it expects to see as those actions play out, a design meant to keep the model's internal sense of "what happens next" grounded in the visual scene.

According to Black Forest Labs' documentation, the model spans several embodiments out of the box: the DROID robot platform (3 cameras, 8 action dimensions), the SO-101 arm (2 cameras, 6 joint commands), video games (256x256 frames, 3-4 control values), and drones (a single onboard camera feed, four movement parameters). DROID and SO-101 ship with prepared checkpoints; the game and drone examples are built by fine-tuning the shared base model with new input and output layers for those controls, using recorded demonstrations.

The release includes the 7B action-base weights, a shared video VAE, and a bundled Qwen3-VL-4B-Instruct text encoder, plus ready-to-use policies for SO-101 and DROID that reference those shared components automatically. Everything ships under the FLUX Kommunity License, with the text encoder separately licensed under Apache 2.0. Black Forest Labs built the release together with NVIDIA, integrating it natively into Hugging Face's LeRobot framework with fine-tuning recipes and a path to edge deployment on NVIDIA Jetson hardware.

The model card is explicit about where responsibility sits once the weights leave the lab: "Nothing in the model bounds joint velocity, force or workspace; the application must enforce those limits," and the license bars using the model or its derivatives "to control a machine in a way that endangers people, without human oversight and a means to stop it."

Context

FLUX 3 Action builds on the same image, video, and audio pretraining as the broader FLUX 3 family, but uses a smaller, deployment-oriented architecture trained specifically to predict actions and future frames together rather than images alone. It arrives as Hugging Face's LeRobot project has become a common integration point for open robot-control models, and as NVIDIA has been pushing Jetson as an edge-inference target for exactly this kind of policy model.

Why it matters

Most widely available open robot-control models have come from robotics-focused labs or research groups; a general-purpose image/video generation company shipping a 7B action model, with checkpoints for both a real robot arm and a research robot platform, plus game and drone variants built on the same base, signals that the underlying "predict the next frame and the next action together" approach is becoming a reusable technique rather than a one-off research result. Bundling fine-tuning recipes and Jetson deployment support alongside the weights also lowers the bar for smaller teams to adapt the model to their own hardware, rather than requiring the kind of large-scale robotics data pipeline that has typically gated this kind of work.