FLUX 3 Action brings video-fast control to robots

FLUX 3 Action is an open-weight, 7-billion-parameter world-action model released by Black Forest Labs, designed specifically for robot control and real-time environment interaction.

Until now, embodied AI developers faced a frustrating tradeoff: run heavy World Action Models like Cosmos 3 Nano that predict video alongside actions for higher accuracy at the cost of massive latency, or pick fast Vision Language Action models like Pi0.5 and sacrifice nearly a quarter of your success rate. FLUX 3 Action eliminates that compromise.

Despite having less than half the parameters of Cosmos 3 Nano (7B vs 16B), FLUX 3 Action takes the top spot on the RoboLab benchmark. Its single-step checkpoint hits a 38.3% success rate, while the guidance-distilled checkpoint reaches 42.2% (topping out at 42.92% on the benchmark table), handily beating Cosmos 3 Nano's 36.8% and Pi0.5's 28.0%.

How it stays fast without sacrificing accuracy

The model is built on top of the multimodal FLUX 3 backbone using Self-Flow pretraining across images, audio, and large-scale video, where video tokens accounted for over 95% of training. Pretraining on raw video is what gives the model an intuitive grasp of physics. Without it, the model scored under 1% on task benchmarks; with it, performance immediately jumped to 11.6%.

During midtraining, Black Forest Labs introduced aligned action data across 14 embodiments, handheld grippers, egocentric video, and 19.55% gameplay recordings. That gaming data means FLUX 3 Action can handle digital screen navigation and computer-use tasks just as well as physical grippers.

Speed comes down to architectural efficiency and distillation. By removing guidance passes and distilling sampling down to a single step, FLUX 3 Action runs 1.52x to 3.95x faster in FP8 than Cosmos 3 Nano across consumer, workstation, and datacenter GPUs. It also predicts a longer action horizon of 2.13 seconds of motion, compared to just 1.0 second for Pi0.5.

Hardware specs and hybrid deployments

The weights are fully available to download alongside documentation on how to fine-tune the model for custom hardware. At 7B parameters, the base checkpoint is compact, taking up only about 14 GB of storage, making it practical to deploy on local workstation GPUs.

In practice, you supply a text goal, live camera feeds, and robot joint positions. The model outputs motor commands together with visual predictions of what the scene should look like next, letting the arm re-evaluate and self-correct on the fly.

The team also tested a hybrid policy using GPT 6 Astra as a high-level planner delegating physical execution to local FLUX 3 Action instances. Running pure GPT 6 Astra on full reasoning costs an average of $13.47 and 16 minutes per success. Pairing it with FLUX 3 Action dropped the cost to $8.77 and execution time to 8 minutes while maintaining a 90% success rate, delivering 53.64% more successes per dollar.

Why it matters

If you are building robot controllers, automation rigs, or latency-sensitive game agents, FLUX 3 Action is the obvious open-weight model to test right now. You get the world-understanding perks of a full world-action model without needing a multi-GPU server rack just to execute a single grip.

More broadly, this release cements a clear design pattern for embodied AI: let expensive frontier models handle reasoning from the cloud, but keep physical motor control local, fast, and lightweight.