Flux 3 Action

FLUX 3 Action (F3A) is a 7B-parameter world action model (WAM) from Black Forest Labs. It takes goals, camera views and joint positions, and jointly predicts future video and robot actions. It is built on FLUX's multimodal Self-Flow pretraining across images, video and audio, then adapted through joint video-action training and finetuning for a target embodiment and action space.

According to the vendor's report, the single-step checkpoint scores 38.3% on the RoboLab benchmark versus 36.8% for Cosmos 3 Nano and 28.0% for Pi0.5. A guidance-distilled checkpoint reaches 42.2%. The vendor also reports 1.34x–2.28x faster real-time factor than Pi0.5 on workstation and datacenter GPUs, with a longer prediction horizon (2.13s vs 1.0s). It also reports hybrid setups, where a frontier reasoning model delegates control to F3A, improving success per dollar.

The weights are published together with a technical report covering pretraining, midtraining, finetuning and inference optimization. Besides robotics, the vendor explores gaming and other latency-sensitive visual agents as testbeds. Benchmark figures are vendor-reported, and the vendor notes that some tasks remain unsolved by F3A alone.