Black Forest Labs Releases FLUX 3 Action: 7B World-Action Model, Weights on Hugging Face
According to a September 24, 2026 report by The Decoder, Black Forest Labs released FLUX 3 Action, an open-source robotics model built on the multimodal FLUX 3, which takes multi-camera video streams as input to predict the next actions and environmental changes; the 7B-parameter model set a success-rate record on RoboLab-120 and runs up to 3.95x faster than the previous best open-source model.
What changed
According to a September 24, 2026 report by The Decoder, Black Forest Labs (BFL) released FLUX 3 Action, an open-source AI model aimed at robotics. It is built on the multimodal model FLUX 3, whose training data is mainly video but also includes images and audio. As a "world-action model," FLUX 3 Action takes multi-camera video streams from a robot's workspace and predicts the next action the agent should take as well as how the environment will change as a result. The report cites BFL as saying that on the RoboLab-120 benchmark, FLUX 3 Action set a success-rate record with only 7 billion parameters — less than half the size of the previous best open-source model — while running up to 3.95x faster than it. BFL says that although large reasoning models have strong planning capabilities, they are often too slow and too bulky for robots, and that this efficiency is critical for on-device deployment. The report also mentions that BFL is bullish on the potential of digital environments, with video games serving as a testing ground for navigation capabilities, and that fast-reacting agents capable of operating computers may emerge later. Model weights are available on Hugging Face. Note: the report date is September 24, 2026, and the above success rate, speed multiple and parameter count are BFL's own claims, based on its own account rather than third-party reproduction.
What it means for developers
The direct implication for developers lies in the model's shape: FLUX 3 Action's input is multi-camera video streams rather than text prompts, and its output is action forecasts and environment-change predictions, making it closer to how robotic policy networks are deployed than to a general conversational model. The 7B parameter range means that low-latency inference on a single GPU or edge device is theoretically more feasible, and BFL's efficiency claim is aimed precisely at the pain point that "large reasoning models are too slow for planning" — if that claim holds under real working conditions, the real-time pressure of on-device closed-loop control would be noticeably lower than calling a remote large model. Weights are provided on Hugging Face, lowering the barrier to experimentation, but note: weights being downloadable does not equal a license permitting commercial use; the report does not state the specific license terms, training-data composition or usage restrictions. The success rate and speed multiple on RoboLab-120 are BFL's own account, lacking independent reproduction, and a leaderboard result does not equal generalization ability on real robots. In addition, the full version of FLUX 3 Action is not currently listed in the KimiSeek live catalog, so its availability, pricing and rate-limit information should be based on BFL and Hugging Face channels — the model has not been onboarded on KimiSeek.
Integration and evaluation
The following are conditional suggestions, not a restatement of the reported facts. First, if a team already has a robotic-arm or mobile-robot pipeline based on VLA or multimodal policies, it is advisable to first record a short trajectory with its own camera setup and run an inference-latency test: run inference on the target hardware with the same number of cameras, resolution and frame rate as production, measure end-to-end action-output latency and VRAM usage, and use "can an action be produced within the control cycle" as the admission criterion, rather than directly trusting the speed multiple in the report. Second, if a team wants to use a world-action model for digital-environment preliminary research (the report mentions game navigation), it should first verify how multi-camera input aligns with the action space: confirm whether the model's output action representation is compatible with the action dimensions of one's own simulator or game interface, compare the model's success rate against the existing scripted policy on a small number of levels with a fixed random seed, and use "whether the success rate is significantly higher than the baseline" as a precondition for further investment. Third, licensing and deployment constraints should be an independent upfront check: before investing engineering effort, confirm whether the weights license permits commercial use and redistribution; if not, it can only be used for internal preliminary research, which directly changes the judgment of whether it is worth integrating into a product pipeline.
Limits and open questions
The report does not give the license name, training-data sources, model input resolution or action-space definition, nor does it state whether inference code, fine-tuning scripts or quantized versions are provided; the fact that weights are available on Hugging Face is not sufficient to infer commercial-use or redistribution rights. The success-rate record on RoboLab-120, the parameter comparison and the 3.95x speed are all BFL's own account, with no third-party reproduction yet, and a leaderboard result cannot be directly extrapolated to real robot tasks. FLUX 3 Action is not currently listed in the KimiSeek live catalog; the model availability in the report refers to BFL and Hugging Face channels, and if it is later listed on KimiSeek, the site's actual list should prevail.