Black Forest Labs Releases FLUX 3 Action for Open-Weight Robot Control
The 7-billion-parameter model predicts actions and future video together, with public weights and LeRobot recipes but only vendor-run evidence of its control performance.
Edited by Tyronne Panaino
Black Forest Labs released FLUX 3 Action on September 23, 2026, as an open-weight model for robot control and other visually grounded action tasks. The 7-billion-parameter system predicts a sequence of actions together with the future video frames those actions are expected to produce, and the company published weights, code and fine-tuning guidance.
The intended audience is robotics teams that want a reusable policy foundation rather than a closed control service. Black Forest Labs presents benchmark and demonstration results across robot arms, games and a simulated indoor drone, but all performance evidence fetched for this article comes from the developers and collaborators. The release confirms that the artifacts exist; it does not independently establish the model's reliability in unfamiliar physical environments.
A world-action model joins seeing and doing
Many robot policies map observations and instructions directly to motor actions. FLUX 3 Action instead uses a world-action-model design that predicts future visual states and actions in one model. The company says this lets video-heavy multimodal pretraining contribute representations that can later be adapted to a particular robot embodiment and action space.
The published architecture is a diffusion transformer with a frozen video variational autoencoder for frames and a frozen Qwen3-VL-4B encoder for instructions. Black Forest Labs says the model was derived from its multimodal FLUX 3 backbone and then trained further with action-aligned data. The official report says video accounted for more than 95 percent of pretraining tokens, while a later action-focused phase mixed games, egocentric hand data, handheld grippers and teleoperation across multiple embodiments.
That design makes the release different from a vision-language model that only explains what a robot should do. FLUX 3 Action is intended to emit executable control predictions as well as a visual forecast. The robot performs a set of actions, observes the environment again and replans from the new state.
Public artifacts still come with a specific license
Black Forest Labs published model weights under the FLUX Kommunity License v1.0 and linked the implementation code and training recipes. Calling the release open-weight describes access to the model artifacts; it should not be read as a blanket claim that every use is unrestricted. Teams need to evaluate the license and the deployment requirements for their own commercial or research setting.
A Hugging Face organization article describes a LeRobot integration developed with Hugging Face and NVIDIA. In that workflow, FLUX 3 Action becomes a LeRobot policy class with parameter-efficient recipes for adapting it to a task. The article says one demonstrated robot-arm policy was adapted from roughly 200 teleoperated episodes covering related pick-and-place tasks.
The developers also show the policy handling containers that were not in its adaptation set and changes to camera position. Those demonstrations are useful evidence of the design target, but they are curated examples rather than an independent robustness study. They do not establish safe operation around people, reliable recovery across arbitrary hardware or transfer to every unseen scene.
Benchmark leadership is a vendor-reported result
Black Forest Labs reports that its guidance-distilled checkpoint reached about 42.2 percent success on RoboLab, with the accompanying summary table placing FLUX 3 Action at the top of the listed open and closed systems. The report attributes the efficiency trade-off to a smaller backbone and distillation that reduces sampling to a single step.
Benchmark settings, serving precision, hardware and checkpoint choice materially affect those comparisons. The official page presents several variants and reports different success and speed points across them. This article therefore treats the claimed lead as a company-run result, not as a universal ranking or proof that the model will lead on a particular robot fleet.
What builders should verify
The most useful next evidence would be third-party reproduction of RoboLab results, evaluation on physical systems outside the release team, failure analysis for tasks that the policy cannot complete, and clear reporting of latency on commonly available hardware. Safety testing should also examine how the visual forecast behaves when objects move unexpectedly, sensors degrade or a person enters the workspace.
The release is still a meaningful shift: a major image-model developer is making a robot-control model and adaptation path available as downloadable artifacts. Its practical value now depends on whether independent teams can reproduce the reported trade-offs and safely transfer the model beyond the showcased setups.
Status
Confirmed. Internal confidence: Medium. Official Black Forest Labs and organization-controlled Hugging Face pages confirm the model, artifacts and integration, but they trace to the same interested release group and the performance claims lack independent reproduction here.
Sources
- Black Forest Labs — FLUX 3 Action technical report
- Black Forest Labs on Hugging Face — FLUX 3 Action integration and demonstrations
Update note: Last reviewed 2026-09-25. We will revise this post if independent evaluations reproduce the control results or the license and deployment boundaries change.
Sources
Drafted with AI assistance from source briefs; reviewed for citation completeness and label accuracy.