FLUX 3 Action: an open model that plans a robot's next moves

Black Forest Labs has published FLUX 3 Action on Hugging Face, a model that plans a robot's next moves and can be fine-tuned for new robots. The weights are open to download, and commercial use depends on the license terms.

Paylaş
FLUX 3 Action: an open model that plans a robot's next moves

Black Forest Labs published its FLUX 3 Action model on Hugging Face on September 23, 2026. According to the company's announcement, the model plans a robot's next moves and can be fine-tuned for a new robot or task. The weights are open for anyone to download, and commercial use depends on the license terms.

What FLUX 3 Action is

The company calls it a "world action model": it predicts how the scene will change and what the robot should do, together. It takes a camera image and a written instruction and produces the moves for roughly the next two seconds. The robot carries out the first few, looks again and refreshes its plan. The 7-billion-parameter model is derived from the base of FLUX 3, the company's image and video model.

What was released

Three sets of weights were shared: a base version for adapting to new robots, a version for the Franka robot arm in the widely used DROID dataset, and a version for the low-cost SO-101 arm in Hugging Face's LeRobot project. According to the company, the LeRobot integration was built together with NVIDIA and Hugging Face. The company also released examples of the same model playing two video games and flying a drone in simulation. The code for the fine-tuning tools is open under the Apache 2.0 license.

Who can use it

The weights are distributed under the FLUX Kommunity License. According to the license text, non-commercial use is free; commercial use is open only to companies below a set annual revenue, and others need a separate license. Military use and surveillance are prohibited by the license. According to the model card, the model is meant for data center GPUs; the company says it uses about 32 GB of memory on an NVIDIA H200.

According to the company, the model ranks first on the RoboLab leaderboard run by NVIDIA's research team. The model cards also carry a safety warning: the model does not limit joint speed, force or workspace on its own, so the application must set those limits and keep a stop button within reach. The company recommends testing in simulation or with the arm's safety limits on before running the model near people, and notes that it cannot solve every task on its own. For a fuller picture, it is worth following evaluations from independent researchers as well.

Sources: Hugging Face: FLUX 3 Action announcement, Hugging Face: FLUX 3 Action models, GitHub: flux-action, Black Forest Labs: research notes.

Devamını oku

Liquid AI تطلق أكبر نماذج LFM2.5 المفتوحة لتشغيل الوكلاء على الأجهزة

Liquid AI تطلق أكبر نماذج LFM2.5 المفتوحة لتشغيل الوكلاء على الأجهزة

أطلقت Liquid AI نموذج LFM2.5-24B-A2B كنموذج مفتوح، ويضم 24 مليار معامل إجمالا لكنه يشغل نحو 2.5 مليار منها في كل خطوة. وبحسب النتائج التي نشرتها الشركة، يولد النموذج المدرب للوكلاء الذين يستخدمون الأدوات 184 رمزا في الثانية على Apple M5 Max.

GlobalFeed Editor tarafından