Black Forest Labs (BFL), the German AI startup best known for its FLUX image and video models, is pushing deeper into robotics with FLUX 3 Action, a 7-billion-parameter, open-weight “World Action Model” designed to take camera observations, a robot’s current state and natural-language instructions and translate them into physical actions.
The company says FLUX 3 Action reaches a 42.92% overall success rate on NVIDIA’s RoboLab-120 benchmark, which would put it ahead of every model currently listed on the public leaderboard.
More significantly for developers looking to deploy rather than simply benchmark a robotics model, BFL says FLUX 3 Action has 7 billion parameters, compared with 16 billion for NVIDIA’s Cosmos3-Nano-Policy, while running 1.43 times faster — meaning it uses just 44% as many parameters.
BFL is also testing the same architecture outside physical robotics, and says the approach can be scaled to simulations, games, and computer use — environments where a model similarly needs to interpret changing visual state and decide what action to take next.
The company says it trained task-specific FLUX 3 Action policies to fly drones in real life and even successfully beat the hit computer game "Doom" with no in-game deaths, piloting the first-person view and main character's gun arm.
The company characterizes these results as early experiments rather than production capabilities.
BFL told VentureBeat it will be releasing the weights, code, fine-tuning recipe, benchmarks and reproducible examples, including an implementation using the relatively inexpensive SO-101 robot and Hugging Face’s LeRobot framework. Teams can also fine-tune the model on demonstrations collected from their own robots.
BFL first previewed FLUX 3 Action model when it launched the broader FLUX 3 family in July. At that time, Action was accessible only through selected research and commercial partners, while the broader open-weight FLUX 3 Dev model remained promised for later.
BFL claims a new high-water mark on RoboLab
RoboLab is an NVIDIA-developed simulation benchmark intended to test general-purpose robotic policies across 120 tasks, spanning visual understanding, relational reasoning and procedural skills, with multiple difficulty levels and variations in how precisely a task is described. NVIDIA designed it in part to avoid the saturation that can make older robot benchmarks less useful as policies improve.
BFL says FLUX 3 Action’s 42.92% score therefore exceeds Cosmos 3, the previous highest-scoring open model on RoboLab, by 6.1 percentage points. Cosmos3-Nano-Policy has 16 billion parameters compared with FLUX 3 Action’s 7 billion.
But the competitive picture is broader than the RoboLab ranking alone. Immediately before FLUX 3 Action’s release, OASIS WAM held the top overall position on RoboLab-120 at 39.0%. Still, FLUX 3 Action comes out ahead.
The size comparison is meaningful against Cosmos, although FLUX 3 Action is not unusually small across the entire robotics-model market.
Ai2’s MolmoAct 2 is a roughly 5-billion-parameter model, while NVIDIA distributes a 3-billion-parameter GR00T N1.7 checkpoint. The distinction is instead that BFL is claiming a higher RoboLab success rate from substantially fewer parameters than the particular open WAM that previously led that benchmark.
Note that MolmoAct 2 and newer versions of NVIDIA’s GR00T and other robotics models are evaluated on different mixtures of real-world and simulated robotics tasks, and therefore don't currently appear alongside FLUX 3 Action in RoboLab.
That makes direct numerical comparisons difficult — a recurring problem in robotics, where hardware, task distributions, demonstrations and evaluation environments can differ dramatically between research groups.
But BFL is also emphasizing inference speed. Its supplied results compare models using real-time factor, essentially the amount of inference latency divided by the span of real-world robot action produced. Lower numbers are better because they mean the model spends less compute time deciding what to do relative to the amount of robot motion it controls.
BFL says distilled FLUX 3 Action variants establish a new Pareto frontier between RoboLab success rate and inference speed across datacenter GPUs, and that the model surpasses π0.5 in real-time factor while maintaining substantially higher task success.
Those are important deployment metrics: a robot policy that scores well but cannot continuously produce actions fast enough risks pauses, jitter or delayed reactions. But the new FLUX result was supplied by BFL ahead of publication; NVIDIA’s public leaderboard had not yet been updated with FLUX 3 Action when checked before launch.
It is also worth separating the different model categories being compared. BFL and NVIDIA characterize FLUX 3 Action and Cosmos 3 as World Action Models, or WAMs, which connect predictions about how the world will evolve with action generation.
Physical Intelligence’s π0.5 and NVIDIA’s GR00T family are vision-language-action models, or VLAs, while Ai2 describes MolmoAct 2 as an “Action Reasoning Model.” All can compete for overlapping robot-manipulation workloads, but they do not necessarily learn or generate actions in the same way.
It predicts what happens next — and what the robot should do next
The architecture follows the thesis BFL has been building toward for much of this year: that a model trained deeply enough on video should learn useful representations of motion, contact, object behavior and cause-and-effect, and those representations can then be adapted for robotic control.
FLUX 3 Action builds on the image, video and audio pretraining behind FLUX 3, but BFL says it uses a smaller architecture optimized for practical deployment. During additional training, the model learns to predict future video frames and actions together.
At inference time, it takes recent camera frames, the system state and a task description, and returns the next 32 actions along with a prediction of how the visual scene will evolve. The robot then observes the updated environment and repeats the process.
The underlying approach traces back to BFL’s Self-Flow research, which attempts to make a generative model learn useful representations of its inputs at the same time it learns to generate them. VentureBeat previously covered Self-Flow as BFL’s attempt to eliminate dependence on separate frozen representation models such as CLIP or DINO. BFL later scaled that work into FLUX 3.
BFL has released Self-Flow inference code and an ImageNet checkpoint under an Apache 2.0 license, though the full training code and video-model checkpoint have not been made public.
That connection between generation and robotics was already visible in FLUX-mimic, built with Swiss startup mimic robotics. BFL said the earlier system used representations from FLUX 3’s video backbone to control robots performing industrial manipulation tasks and had been tested in Audi production environments.
FLUX 3 Action takes the approach into a model other robotics teams can adapt themselves.
BFL says the policy shown in its new demonstrations was fine-tuned from around 200 teleoperated episodes covering a handful of related pick-and-place tasks. The company says the objects shown in the test videos were not present in that training set.
One supplied demo is particularly notable because the robot initially makes a mistake and then attempts the task again. In an accompanying interview transcript, the human researchers described the behavior this way: “the model failed first, and then it tried again and better.”
That is still a company-selected demonstration rather than evidence that autonomous recovery will generalize across robots and environments, but it illustrates the behavior BFL hopes a pretrained world model can provide without teams explicitly demonstrating every possible failure mode.
BFL is not alone in arguing that a broadly pretrained model can reduce the amount of robot-specific data required. Google DeepMind says its proprietary Gemini Robotics On-Device 2 VLA can adapt to new robot embodiments with fewer than 200 examples, although access remains limited to trusted testers.
On the open side, Ai2’s MolmoAct 2 may be the most directly relevant counterpoint to BFL’s release strategy. Ai2 has released not only MolmoAct 2’s weights but its training code, fine-tuning scripts, datasets, evaluation rollouts and LeRobot integration. Ai2 reports an 87.1% average success rate across five of its own real-world Franka evaluation tasks, versus 45.2% for π0.5 — results that cannot be compared directly with RoboLab’s percentages because the task sets and evaluation methodology differ.
That distinction matters for developers assessing these systems: robotics still lacks a single benchmark that cleanly establishes an overall model leader across simulation, real hardware, different embodiments and different types of manipulation. FLUX 3 Action’s result establishes its position on RoboLab; it does not resolve the wider contest among WAMs, VLAs and action-reasoning models.
BFL moves from image generation toward physical AI
FLUX 3 Action also shows how rapidly BFL’s scope has expanded since the company emerged in 2024.
The startup was founded by researchers behind technologies including latent diffusion and Stable Diffusion and initially built its reputation around the FLUX.1 family of image models. BFL subsequently introduced a commercial API with FLUX1.1 Pro, then expanded its open-core strategy with the FLUX.2 family, including smaller [klein] models designed for high-speed inference and local deployment.
BFL raised a $300 million Series B at a $3.25 billion post-money valuation in December 2025, bringing its total funding to more than $450 million. The company operates from Freiburg and San Francisco.
FLUX 3 Video became generally available through BFL’s API in August, with pay-as-you-go pricing starting at $0.06 per second for Draft HD, $0.17 per second for standard HD and $0.29 per second for FHD text- or image-to-video generation. BFL now also offers video upscaling to higher resolutions through a separate endpoint.
However, the release materials for the new FLUX 3 Action don't yet announce an Action API, API pricing or a specific commercial license for the weights. At press time, BFL’s existing public FLUX 3 page still described Action as rolling out through selected research and commercial partners,. Its public product documentation likewise lists FLUX 3 primarily as a video-generation API rather than exposing a FLUX 3 Action endpoint.
For robotics teams, that makes self-hosting and fine-tuning the central proposition for now.
That approach puts FLUX 3 Action into an increasingly crowded open or downloadable robotics ecosystem:
NVIDIA’s GR00T N1.7, released in April, is an open 3B VLA aimed particularly at generalized humanoid and cross-embodiment manipulation, and it can be post-trained for particular robots, tasks and environments.
MolmoAct 2, released in May, similarly targets local adaptation and is integrated directly into LeRobot.
By contrast, Google’s Gemini Robotics On-Device 2, released in July, represents the proprietary end of the spectrum: it is designed to run locally on robot hardware, but Google currently distributes it only to selected testers rather than releasing its weights publicly.
And that may ultimately be the more consequential part of the release. Rather than requiring a team to train a large robotics foundation model itself, BFL is betting developers can begin with a relatively compact model carrying broad visual and physical representations, collect demonstrations on their own hardware, and adapt it to tasks such as sorting, assembly, packaging, navigation and pick-and-place.
The open-weight release also gives teams a different deployment option from hosted robotics services: they can inspect and modify the model, run inference within their own infrastructure and potentially keep proprietary robot demonstrations and operational data inside their own environment. Exactly how commercially permissive that option will be still depends on BFL publishing the final license terms.
The benchmark numbers suggest that approach can be competitive in simulation. The broader competitive question is now less whether FLUX 3 Action can beat one particular model on one leaderboard than whether its combination of world-model pretraining, 7B size, open weights and fine-tuning workflow proves useful across the messy variety of real robots that developers actually operate.
The next test is whether outside teams can reproduce those gains on their own robots — and whether FLUX 3 Action’s smaller size and open weights make doing so practical enough for production.