The 7-billion-parameter open-weight model claims to outperform larger competitors in robotic manipulation tasks while running on consumer hardware
Black Forest Labs launched FLUX 3 Action, a 7-billion-parameter open-weight robotic model built on top of its newly unveiled FLUX 3 multimodal foundation model, positioning itself as a serious contender in the rapidly heating physical AI space.
The model reportedly outperforms larger competitors on the RoboLab-120 benchmark, though there is no public confirmation yet of a distinct 7-billion-parameter model specifically branded as “FLUX 3 Action,” nor is it listed among the top contenders on the RoboLab-120 leaderboard.
What FLUX 3 actually does
FLUX 3 integrates images, video, and audio into a single foundation model using a technique BFL calls Self-Flow, which applies flow matching to unify these modalities. The underlying architecture pairs a video-prediction backbone with a streamlined decoder that translates visual sequences into robotic control commands.
On the video generation side, FLUX 3 supports clips up to 20 seconds with synchronized audio. In human evaluations, the model was preferred over Runway Gen-4.5 by 77% of evaluators and over Luma Ray 3.2 by 93%.
BFL reports end-to-end latency between 80 and 101 milliseconds on consumer-grade hardware. Fine-tuning can be achieved with as little as 30 minutes of robot demonstration data.
Real-world performance and partnerships
The model has already been tested on soft-body manipulation tasks at Audi, where BFL reports a 95% success rate.
BFL has also initiated a partnership with mimic robotics, a Zurich-based company, to develop what it calls FLUX-mimic. This collaboration represents the first commercial deployment channel for the action capabilities, with select companies getting early access to both FLUX Action and FLUX-mimic functionalities.
BFL has committed to releasing the weights of the FLUX 3 Dev multimodal backbone by the end of 2026, though the video and action capabilities will initially remain gated.
The competitive landscape is getting crowded
FLUX 3 Action enters a market where competing models like OASIS WAM and various iterations of Cosmos3 are also vying for dominance in physical AI benchmarks. The claim of topping the RoboLab-120 leaderboard has not been independently validated.
