Black Forest Labs Open-Sources FLUX 3 Action, a 7B World-Action Model for Robots

Black Forest Labs released FLUX 3 Action on September 23, 2026: an open-weights 7-billion-parameter world-action model (WAM) adapted from its multimodal FLUX 3 backbone. The model takes one or more camera frames, a robot state vector, and a text instruction, then returns the next roughly two seconds of motor actions—32 actions per call—optionally alongside predicted future frames.

Unlike vision-language-action models that output actions alone, FLUX 3 Action jointly denoises video and action tokens in one diffusion transformer. A frozen video VAE encodes frames and a frozen Qwen3-VL-4B encodes the instruction. Weights ship under the FLUX Kommunity License v1.0 in a Hugging Face collection that includes a base package plus DROID and SO-101 policy checkpoints; training and fine-tuning code live in the flux-action GitHub repository.

On the RoboLab-120 leaderboard, Black Forest Labs reports a 42.92% overall success rate for FLUX 3 Action at 7B parameters, ahead of NVIDIA’s open Cosmos3-Nano-Policy at 36.8% with 16B parameters and Physical Intelligence’s π0.5 at 28.0%. The company also claims up to about 3.95× faster inference versus Cosmos 3 Nano depending on GPU and precision, after guidance and step distillation that keep joint video-action prediction intact.

For real hardware, BFL integrated the policy with Hugging Face’s LeRobot stack in collaboration with NVIDIA, including parameter-efficient fine-tuning recipes. Independent third-party testing by Positronic Robotics on a Franka arm scored FLUX 3 Action 28 of 30 attempts across ten DROID-style tasks, compared with 27/30 for Cosmos 3 Nano, 20/30 for DreamZero, and 13/30 for π0.5 under matched conditions. The same fine-tuning recipe was also demonstrated on SO-101 pick-and-place, indoor drone control in Isaac Sim, and small game environments used as fast policy testbeds.

DigiEditorial verified the release against Black Forest Labs’ primary research page and the Hugging Face announcement blog, with corroborating coverage from The Decoder; this article summarizes publicly confirmed details only.