38.3 percent success rate with just 7 billion parameters – beating a competing model with more than twice as many parameters. That number alone explains why FLUX 3 Action has been the talk of developer forums, GitHub, and robotics labs since its release day. Black Forest Labs, the AI company based in Freiburg, Germany and San Francisco, unveiled FLUX 3 Action on September 24, 2026: an open World Action Model that controls robot arms directly from camera images, sensor data, and text commands. In a field long dominated by huge, closed models, this is a remarkable counter-proposal: smaller, faster, open – and, according to an independent benchmark, still at the top.
The following article breaks down what FLUX 3 Action actually does, how it differs from classic Vision-Language-Action models, which numbers on the RoboLab-120 leaderboard really matter, and how developers can integrate the model into their own projects.
Table of Contents
Key Takeaways
- FLUX 3 Action is an open 7-billion-parameter World Action Model from Black Forest Labs, released on September 24, 2026, built on the multimodal FLUX 3 backbone.
- On Nvidia Research’s RoboLab-120 leaderboard, the model reaches a 38.3 percent success rate (distilled variant up to 42.92 percent), beating Cosmos3-Nano-Policy (16B, 36.8 percent) and Pi0.5 (3.3B, 28.0 percent).
- Despite being less than half the size of Cosmos3-Nano, FLUX 3 Action runs up to 3.95x faster on B200 GPUs.
- Weights and code are openly available (Hugging Face, GitHub) under the FLUX Kommunity License v.1.0; the Qwen3-VL-4B-Instruct text encoder is under Apache-2.0.
- Combined with a reasoning model like GPT 6 Astra as a hybrid policy, the success rate climbs to 90 percent according to BFL’s own report, at 29 percent lower cost per successful task.
What Is a World Action Model – And How Is It Different from a VLA?
To place FLUX 3 Action correctly, it helps to look at the difference between two model classes that are frequently lumped together in robotics AI.
A classic Vision-Language-Action model (VLA) takes a camera image, a text instruction, and the robot’s current state, and directly predicts the next action from that – for example, how a gripper arm should move. This works well, but has a weakness: the model does not really “understand” how the world will change as a result of its action. It reacts without thinking ahead.
A World Action Model (WAM) like FLUX 3 Action goes a step further. It not only predicts the next movement, but simultaneously predicts what the scene will look like afterward – i.e., the resulting video frames [1]. The model essentially builds an internal model of the physical world and simulates the consequences of its own actions before or while acting. This is conceptually closer to a kind of “mental preview” than to a pure stimulus-response chain.
In practice, this means: FLUX 3 Action processes images from multiple cameras simultaneously, the robot’s current state (joint angles, gripper position, and similar data), and a natural-language instruction. From these three sources of information, it jointly computes the next motion actions and a prediction of the upcoming video frames [1][2]. This joint prediction of action and visual consequence is the central technical difference from a pure VLA.
FLUX 3 Action in Detail: Architecture and Origins
FLUX 3 Action is not built on an entirely new base architecture; instead, it uses the multimodal FLUX 3 backbone that Black Forest Labs had already introduced in July 2026 [3][4]. This backbone was trained from the ground up for image, video, and audio processing, giving it a correspondingly deep understanding of visual and temporal relationships – a basic prerequisite for a model like FLUX 3 Action to predict plausible future video frames at all.
On top of this video-heavy foundation, Black Forest Labs adds the robotics component: a compact but capable 7-billion-parameter model specifically trained to control real robot arms [1][2]. Qwen3-VL-4B-Instruct is used as the text encoder, which is separately available under the permissive Apache-2.0 license [8].
“FLUX 3 Action shows that a smaller, purpose-trained model can beat a significantly larger competing system in a tough, independently measured benchmark.” – paraphrased characterization of the release by independent trade media [6]
Performance on the RoboLab-120 Leaderboard
The real proof of FLUX 3 Action’s capability comes from the RoboLab-120 leaderboard, which is run by Nvidia Research – importantly, Nvidia merely provides the independent test environment here and is not the developer of FLUX 3 Action [5][6].
The table below summarizes the key comparison figures:
| Model | Developer | Parameters | RoboLab-120 Success Rate |
|---|---|---|---|
| FLUX 3 Action (single-step) | Black Forest Labs | 7B | 38.3 percent |
| FLUX 3 Action (Distilled) | Black Forest Labs | 7B | up to 42.2/42.92 percent |
| Cosmos3-Nano-Policy | Nvidia | 16B | 36.8 percent |
| Pi0.5 | Third party (open VLA) | 3.3B | 28.0 percent |
What stands out isn’t just the raw success rate, but the ratio of performance to size: FLUX 3 Action is less than half the size of Cosmos3-Nano-Policy, yet clearly outperforms it [5][6]. On Nvidia B200 GPUs, FLUX 3 Action also runs up to 3.95x faster when both models are run at FP8 precision [5][6]. For production robotics applications, where latency directly determines the physical system’s reaction speed, this speed advantage is practically at least as important as the raw success rate.
Hybrid Approach: FLUX 3 Action Plus a Reasoning Model
Black Forest Labs shows in its own report that FLUX 3 Action’s results can be improved further by combining the model with a separate reasoning model – specifically GPT 6 Astra acting as the planning instance [5][6].
In this hybrid setup, the reasoning model handles higher-level task planning, while FLUX 3 Action delivers fast, fine-motor execution. According to BFL’s report, the result is:
- 90 percent of all test tasks are solved successfully.
- Cost per successful task drops to $8.77 at around 8 minutes of execution time.
- For comparison: pure reasoning without the fast action policy costs $13.47 and takes around 14 minutes.
- That amounts to 29 percent lower cost and 40 percent shorter time per successful task.
This hybrid approach points to a broader trend: slow, expensive reasoning models for planning, fast, specialized action models for execution. FLUX 3 Action is explicitly designed for this role as a fast, precise execution layer.
Open Weights, License, and Availability
A central feature of FLUX 3 Action is its openness. The weights are available on Hugging Face in several variants, including base, so101, and droid repositories under the black-forest-labs/flux-3-action namespace [8]. The corresponding code is available on GitHub in the black-forest-labs/flux-action repository.
Legally, FLUX 3 Action is released under the FLUX Kommunity License v.1.0. The separately usable Qwen3-VL-4B-Instruct text encoder, by contrast, is under the more permissive Apache-2.0 license. This split lets developers reuse the language-understanding component relatively freely, while the actual action component remains subject to the Kommunity License’s terms [8].
For practical integration, it’s relevant that FLUX 3 Action, as a compact 7B model, is designed to run – with quantization – on significantly smaller hardware than high-end data-center GPUs, an aspect that makes it attractive for research teams and smaller robotics companies without their own GPU clusters.
Safety Notes: What FLUX 3 Action Does Not Handle Itself
Black Forest Labs explicitly points out an important limitation in its documentation: FLUX 3 Action outputs only joint targets – that is, target positions for the robot’s joints. The model itself does not limit speed, force, or the robot’s permissible workspace [1].
In practice, this means: any application using FLUX 3 Action must implement its own safety limits and a working emergency-stop mechanism. The model makes no independent safety decisions and, under its license terms, must not be used to endanger people without human oversight. For developers integrating FLUX 3 Action into real hardware, this point is not a footnote – it’s a hard technical and legal requirement for the surrounding system architecture.
Product Context: FLUX-mimic as a Predecessor
FLUX 3 Action does not stand in isolation within Black Forest Labs’ portfolio. Back in July 2026, the company, together with robotics partner mimic robotics, introduced the FLUX-mimic model. That system is being tested in production at Audi, among other places, specifically for manipulating soft, deformable objects – a task that traditionally poses major problems for classic, rigid grasping algorithms.
FLUX 3 Action extends this line with a more general, more broadly applicable World Action Model that isn’t tailored to a specific material class, but instead serves as a versatile foundation for a wide range of robot platforms – from research arms to industrial applications.
Conclusion: What FLUX 3 Action Means for Robotics Development
FLUX 3 Action marks a point at which open robotics AI can compete with closed and significantly larger systems. With 7 billion parameters, the model achieves a higher success rate on the independent RoboLab-120 leaderboard than Nvidia’s Cosmos3-Nano-Policy with 16 billion parameters – and does so at up to 3.95x higher speed on comparable hardware. The combination of open weights, traceable code, and a clearly documented license makes FLUX 3 Action an attractive starting point for research teams, startups, and established robotics companies alike.
For developers who want to get started, the following next steps are recommended: for a broader overview of how today’s leading AI language models stack up, see our AI Model Leaderboard 2026. First, check the model cards and repositories on Hugging Face (base, so101, and droid variants) to identify the right configuration for your own robot platform. Next, take a look at the black-forest-labs/flux-action GitHub repository to plan integration into existing control pipelines. Anyone deploying FLUX 3 Action on real hardware should implement their own safety limits for speed, force, and workspace from the outset, since the model explicitly does not handle this itself. And anyone aiming for maximum success rates should consider the hybrid approach of FLUX 3 Action plus a reasoning model like GPT 6 Astra, to sensibly separate planning from execution.
References
[1] Flux3 Action Overview – https://docs.bfl.ai/flux_3/flux3_action_overview [2] Flux 3 Action – https://bfl.ai/models/flux-3-action [3] Flux3 Overview – https://docs.bfl.ai/flux_3/flux3_overview [4] Flux 3 – https://bfl.ai/models/flux-3 [5] Black Forest Labs Releases Flux 3 Action A 7b Open Weights World Action Model That Tops Robolab 120 – https://www.marktechpost.com/2026/09/24/black-forest-labs-releases-flux-3-action-a-7b-open-weights-world-action-model-that-tops-robolab-120/ [6] Black Forest Labs Debuts Flux 3 Action An Open Weights Ai Robotics Model That Tops The Leaderboard At Half The Size Of Its Competition – https://venturebeat.com/infrastructure/black-forest-labs-debuts-flux-3-action-an-open-weights-ai-robotics-model-that-tops-the-leaderboard-at-half-the-size-of-its-competition [7] Models – https://bfl.ai/models [8] Flux 3 Action – https://huggingface.co/collections/black-forest-labs/flux-3-action [9] Blog – https://bfl.ai/blog [10] Black Forest Labs Open Weights Flux 3 Action A 7b Robotics Model – https://aiweekly.co/alerts/black-forest-labs-open-weights-flux-3-action-a-7b-robotics-model
