Black Forest Labs' FLUX 3 Action Tops RoboLab at Half the Size

Black Forest Labs' 7B-parameter open-weight FLUX 3 Action claims a 42.92% RoboLab-120 success rate, beating NVIDIA's 16B Cosmos3-Nano-Policy.

Black Forest Labs (BFL) unveiled FLUX 3 Action, a 7-billion-parameter open-weight robotics model built on the FLUX 3 multimodal family it launched in July, VentureBeat reports. The model takes recent camera frames, the robot's state and a task description, and returns the next 32 actions along with a prediction of how the visual scene will evolve. BFL says it will release the weights, code, fine-tuning recipe, benchmarks and reproducible examples, including an implementation using the inexpensive SO-101 robot and Hugging Face's LeRobot framework.

A company best known for image and video generation is now claiming the top spot on a robotics benchmark. According to VentureBeat, BFL says FLUX 3 Action reaches a 42.92% overall success rate on NVIDIA's RoboLab-120 benchmark, which would place it ahead of every model on the public leaderboard — beating NVIDIA's Cosmos3-Nano-Policy despite using just 44% as many parameters (7B versus 16B) while running 1.43 times faster. Crypto Briefing cautions that the leaderboard had not yet been updated with the model at publication, so the top-spot claim is not independently validated.

RoboLab-120 spans 120 tasks across visual understanding, relational reasoning and procedural skills. VentureBeat reports that BFL's score of 42.92% exceeds Cosmos 3, the previous best open model, by 6.1 percentage points, and tops OASIS WAM's 39.0%. BFL also says distilled variants set a new Pareto frontier between success rate and inference speed, surpassing π0.5 in real-time factor. The company showed task-specific policies flying drones in real life and beating 'Doom' with no in-game deaths, characterizing these as early experiments. Crypto Briefing adds that BFL reports 95% success on soft-body manipulation tasks at Audi and fine-tuning with as little as 30 minutes of demonstration data.

VentureBeat notes the RoboLab results were supplied by BFL ahead of publication and NVIDIA's leaderboard had not yet listed the model. Benchmark comparisons in robotics are also messy: Ai2's MolmoAct 2 (about 5B parameters) and NVIDIA's GR00T N1.7 (3B) are evaluated on different task mixtures and don't appear alongside FLUX 3 Action on RoboLab. The field spans different model categories too — World Action Models like FLUX 3 Action and Cosmos 3, vision-language-action models like π0.5 and GR00T, and Ai2's 'Action Reasoning Model' framing for MolmoAct 2 — so no single benchmark crowns an overall leader.

BFL is betting developers can start with a compact model carrying broad visual and physical representations, collect demonstrations on their own hardware, and adapt it — VentureBeat notes the showcased policy was fine-tuned from roughly 200 teleoperated episodes, with objects in the test videos absent from training. BFL raised a $300 million Series B at a $3.25 billion valuation in December 2025. For now there is no Action API or commercial license announced; self-hosting and fine-tuning are the central proposition. Crypto Briefing reports BFL has committed to releasing the FLUX 3 Dev backbone weights by the end of 2026, with video and action capabilities initially gated.