무슨 일이 일어났나요?
Black Forest Labs (BFL) released FLUX 3 Action, a 7-billion-parameter open-weight 'World Action Model' designed for robotics. The model takes camera observations, robot state, and natural language instructions to generate physical actions. BFL reports a 42.92% success rate on NVIDIA's RoboLab-120 benchmark, surpassing the previous leader, NVIDIA's Cosmos3-Nano-Policy (16B parameters), by 6.1 percentage points. The release includes weights, code, and recipes, allowing developers to adapt the model to specific robots using frameworks like LeRobot.
Black Forest Labs, known for its FLUX image and video models, has expanded into robotics with the release of FLUX 3 Action. This 7-billion-parameter model is an open-weight 'World Action Model' (WAM) that processes camera inputs, system states, and text instructions to output a sequence of 32 physical actions along with a prediction of the visual scene's evolution.
According to VentureBeat, BFL reports that FLUX 3 Action achieves a 42.92% overall success rate on NVIDIA's RoboLab-120 benchmark. This score exceeds the previous top open model, NVIDIA's Cosmos3-Nano-Policy, by 6.1 percentage points. Notably, FLUX 3 Action uses only 7 billion parameters compared to Cosmos's 16 billion, and BFL claims it runs 1.43 times faster, establishing a new Pareto frontier for efficiency and performance.
The release includes the model weights, code, recipes, and reproducible examples. BFL states that teams can fine-tune the model using demonstrations collected from their own robots, specifically mentioning compatibility with the SO-101 robot and Hugging Face’s LeRobot framework. The company also demonstrated the model's capabilities in non-robotic environments, including flying drones and playing the video game Doom without in-game deaths, though these are characterized as early experiments.
BFL first previewed FLUX 3 Action in July when launching the broader FLUX 3 family, but it was initially restricted to selected research and commercial partners. The current release marks a shift toward broader accessibility, although specific commercial API pricing for the Action model has not yet been announced. The model builds on BFL's Self-Flow research, which aims to learn useful representations of motion and cause-and-effect from video data without relying on separate frozen representation models.
왜 중요한가요?
This release is significant because it provides a compact, open-weight alternative to larger proprietary or closed robotics models. By achieving higher benchmark scores with fewer parameters (7B vs 16B) and faster inference, BFL lowers the barrier to entry for teams wanting to deploy advanced robotic policies without massive compute resources. The open-weight nature allows for local and data privacy, which is critical for industrial applications where proprietary data cannot leave the facility. It also validates the 'World Action Model' approach, where video informs physical control, potentially reducing the need for extensive robot-specific data collection.
The release of FLUX 3 Action is consequential for the robotics industry because it offers a high-performance, open-weight option that challenges larger, more resource-intensive models. By achieving superior benchmark results with fewer parameters, BFL demonstrates that efficiency and performance are not mutually exclusive, which is critical for edge deployment where compute resources are limited.
For developers, the open-weight nature of the model is a key differentiator. It allows teams to inspect, modify, and fine-tune the model within their own infrastructure, ensuring that proprietary robot demonstrations and operational data remain private. This is a significant advantage over hosted robotics services or proprietary models like Google's Gemini Robotics On-Device 2, which are limited to selected testers.
The model's architecture, which combines video with action generation, aligns with the growing trend of using generative models to understand physical dynamics. This approach may reduce the amount of robot-specific data required for training, as the model can leverage broad visual and physical representations learned from video. This could accelerate the development of robotic systems for tasks such as sorting, assembly, and navigation.
However, the practical impact depends on the final license terms and the ability of outside teams to reproduce the reported results. Robotics benchmarks are often fragmented, and direct comparisons between different model architectures (WAMs, VLAs, and action-reasoning models) are difficult. The success of FLUX 3 Action will ultimately be determined by its performance in real-world, diverse robotic environments rather than just simulation benchmarks.
대화형 메커니즘: 실제로 작동하는 방식
이 개발의 이면에 있는 기본 기술을 대화식으로 살펴보세요.
In AI, what are a model's "parameters"?
다음에 무엇을 볼 것인가
Developers should monitor the availability of the final commercial license terms, as the current release is primarily for research and partner access. Independent reproduction of the benchmark results on diverse hardware is crucial, as robotics benchmarks often lack standardization. Additionally, watch for the integration of FLUX 3 Action into broader commercial APIs or whether BFL expands its open-core strategy to include more robotics-specific endpoints.
The most immediate factor to watch is the publication of the final commercial license terms for FLUX 3 Action. While BFL has released the weights and code, the specific conditions for commercial use, redistribution, and modification are not yet fully detailed. This will determine how widely the model can be adopted in industrial settings.
Independent verification of the benchmark results is crucial. BFL's reported 42.92% success rate on RoboLab-120 is based on their own testing, and NVIDIA's public leaderboard had not been updated with these results at the time of the report. Reproduction by third-party research groups will be necessary to confirm the model's performance and generalizability.
Developers should also monitor the integration of FLUX 3 Action into existing robotics frameworks and hardware. BFL has highlighted compatibility with the SO-101 robot and LeRobot, but the model's performance on other common robotic platforms, such as those from NVIDIA or Ai2, remains to be seen. The ease of on diverse hardware will be a key determinant of its adoption.
Finally, the competitive landscape in open-weight robotics is evolving rapidly. With models like Ai2's MolmoAct 2 and NVIDIA's GR00T N1.7 already available, the success of FLUX 3 Action will depend on its ability to offer distinct advantages in terms of performance, efficiency, and ease of use. The next few months will likely see more comparisons and benchmarks as the community tests the model in various applications.