Black Forest Labs' FLUX 3 Action tops RoboLab with a 7B open robot model

Black Forest Labs, the lab behind the FLUX image models, has moved into robotics. It has released FLUX 3 Action, an open-weights model with 7 billion parameters designed to control robots, MarkTechPost, The Decoder and GIGAZINE report. The weights are available on Hugging Face.

FLUX 3 Action is what the company calls a World Action Model. Such a model looks at camera images, the robot's state and a text instruction, then predicts at the same time what the scene will look like in the next moments and what actions the robot should take. It is derived from FLUX 3, the lab's multimodal model trained mainly on video, according to The Decoder.

What the benchmark results show

On the RoboLab-120 leaderboard, which groups 120 tabletop tasks simulated in Isaac Sim with ten trials each on a Franka robot arm, FLUX 3 Action ranks first with a 42.92 per cent task success rate, MarkTechPost reports. That is 6.1 percentage points ahead of NVIDIA's Cosmos 3 Nano, the best open model until now at 36.8 per cent, with 56 per cent fewer parameters. The lab's own average over several runs for its recommended checkpoint is 42.24 per cent, plus or minus 0.36.

The results also hold on real hardware, according to the same outlet. Positronic Robotics ran a blind evaluation on a Franka arm with ten tasks and three attempts each: FLUX 3 Action completed 28 of 30 attempts, Cosmos 3 Nano 27, DreamZero 20 and the fast π0.5 policy 13.

• RoboLab-120 (simulation): 42.92 per cent for FLUX 3 Action, 36.8 per cent for Cosmos 3 Nano, 28.0 per cent for π0.5.
• Real robot, blind test by Positronic Robotics: 28 of 30 attempts for FLUX 3 Action, 27 for Cosmos 3 Nano, 20 for DreamZero, 13 for π0.5.
• Speed: up to 3.95 times faster than Cosmos 3 Nano in FP8, according to Black Forest Labs.
• Each call produces 32 actions at 15 Hz, or 2.13 seconds of movement, against 1 second per call for π0.5.

The trade-off it tries to solve

Open robot policies usually force a choice, MarkTechPost explains. Models that predict video, like Cosmos 3 Nano, are more accurate but expensive: on an NVIDIA B200 chip, Black Forest Labs measured that Cosmos 3 Nano needs about 4.7 times more processing time than π0.5 per second of robot motion. Vision-language-action models such as π0.5 are fast but less successful. FLUX 3 Action keeps joint video and action prediction and closes the speed gap with a smaller backbone and distillation.

The model is shipped in three versions, each in two numeric precisions: a base version, a version that runs 1.8 to 2 times faster with slightly better scores, and a single-step version that runs 3.15 to 4 times faster at the cost of 3.5 to 4.3 points of success. On an RTX 5090 consumer graphics card, the fastest version remains slower than π0.5, MarkTechPost notes.

How it was trained

Video accounted for more than 95 per cent of the tokens used to pre-train the FLUX 3 backbone. The action training then mixed that data with video aligned with actions, drawn from game recordings, first-person videos of human hands, handheld grippers and remote-controlled robots across 14 different machines, according to MarkTechPost. Pre-training matters: trained only on robot data, the model stayed below 1 per cent on RoboLab; with pre-training, the same protocol reached 11.6 per cent before further training.

Black Forest Labs also tested the model as the fast half of a hybrid system with OpenAI's GPT-6 Astra acting as the planner, which can accept, edit or replace the predicted actions. The combination solved 90 per cent of episodes at 8.77 dollars and about 8 minutes per success, while Astra alone at maximum effort solved all episodes but cost 13.47 dollars and more than 16 minutes per success, the outlet reports.

Conditions of use

Teams can adapt the model to their own robots. The lab published training recipes, including one showing that a pick-and-place skill for the low-cost SO-101 arm can be learned from about 200 demonstrations, and integrated the model with NVIDIA into Hugging Face's LeRobot platform, with support for NVIDIA Jetson boards, according to MarkTechPost and GIGAZINE. The main policy needs about 32 GB of graphics memory, or fits on 24 GB cards with reduced precision.

Two limits matter. The FLUX Kommunity License only allows non-commercial use. And the model outputs joint targets without built-in limits on speed, force or workspace: the application must enforce them, MarkTechPost stresses. Black Forest Labs also sees uses beyond robotics, such as video games, vehicle control and agents that operate computers, according to The Decoder and GIGAZINE.

What this changes for companies

For US robotics start-ups, integrators and university labs, FLUX 3 Action is a new free baseline to test against NVIDIA's models on their own tasks, with published recipes that lower the cost of trying. It is not a product to deploy on a production line: the licence forbids commercial use without another agreement, and safety limits on force and speed remain the integrator's responsibility. Teams that plan to use it should budget for that engineering work and for a commercial licence discussion before any customer deployment.