Black Forest Labs Releases FLUX 3 Action: Open 7B Robot Model Tops RoboLab-120 Benchmark

Black Forest Labs released FLUX 3 Action on September 22, making the weights freely available on Hugging Face — the first open world-action model from a company outside a major hardware or chip laboratory to exceed the previous best score on NVIDIA's RoboLab-120 manipulation benchmark. The seven-billion-parameter model achieves 42.24% task success in its guidance-distilled configuration, against 36.8% for Cosmos3-Nano-Policy, while running between 1.52 and 3.95 times faster across datacenter and workstation GPUs. More significant than the benchmark position is what the result implies about how to build robot policies: a video-generation backbone that has never seen a robot arm can, through large-scale video pretraining and targeted fine-tuning, outperform models built specifically for physical AI.

How a Video Model Learned to Control Robots

FLUX 3 Action belongs to a class called world-action models, or WAMs — a term introduced in the DreamZero research paper published in February 2026 — to describe systems that jointly predict future video frames and the robot actions needed to produce them in a single model, rather than treating visual understanding and motor control as separate problems. The contrast is with vision-language-action models, or VLAs, such as Physical Intelligence's pi0.5. A VLA uses a pretrained language-vision backbone and outputs actions directly, often skipping video prediction at inference. VLAs are faster and smaller but tend to score lower: pi0.5, at 3.3 billion parameters, achieves 28.0% on RoboLab-120.

WAMs are more accurate because video pretraining forces a model to learn physical causality — that an object pushed from the left will move right, that a cup lifted quickly may tip, that soft materials deform under load. That understanding transfers when the model is fine-tuned to predict not only future frames but also the motor commands that would produce them. The problem is that predicting video and actions jointly produces long token sequences, making WAMs expensive. According to BFL's technical report, Cosmos3-Nano — which uses a roughly 16-billion-parameter backbone — requires approximately 4.7 times the processing time per second of robot motion compared with pi0.5 on the same hardware.

BFL's FLUX 3 Action closes that gap. The model's backbone comes from FLUX 3, announced in July 2026 as a unified multimodal system trained on images, video, and audio using a technique BFL calls Self-Flow — a flow-matching approach that optimizes both generation quality and cross-modal representation in one architecture, rather than aligning separately trained specialist models at inference. Video accounts for over 95% of the training tokens, which is why the model's representations of physical dynamics are strong enough to support manipulation tasks even before any robot data is added.

Architecture, Training, and Inference Optimization

FLUX 3 Action passes multi-camera video feeds, a text instruction, and the robot's current joint positions through the FLUX 3 backbone, then decodes from its output both predicted future video frames and a predicted action sequence. Each forward pass produces 32 motor commands at 15 Hz — 2.13 seconds of robot motion — after which the robot executes those actions, the cameras update, and the model replans. BFL's technical report describes this as a "look, think, move" cycle: the model re-observes the scene after each chunk and adjusts accordingly.

The midtraining phase, which precedes robotics fine-tuning, mixes pretraining video with action data from four sources: gaming recordings (19.55% of samples), egocentric video with hand-pose annotations (13.54%), handheld gripper data (14.03%), and teleoperation across fourteen robot embodiments (15.93%). Gaming footage is represented with a 64-dimensional action space encoding mouse axes, mouse buttons, and keyboard keys — a design choice BFL describes as a deliberate foundation for extending action prediction to computer-use agents, not just physical robots.

Fine-tuning on the DROID robotic dataset using a batch size of 2,048 and joint action space (absolute joint positions rather than end-effector delta) reaches approximately 40% on RoboLab-120 before BFL applies two inference-time optimizations that push performance and speed further. The first, guidance distillation, trains a student model to replicate the guided prediction of the full model in a single forward pass, eliminating the second, unguided forward pass that classifier-free guidance normally requires. This removes a 2x latency penalty with no accuracy loss — the guidance-distilled checkpoint actually improves performance by 0.6 to 1.08 percentage points. The second, step distillation, reduces the number of sampling steps from four to one, providing an additional 3 to 4x speedup. Step distillation does cost 3.51 to 4.32 percentage points of accuracy, but the resulting checkpoint still beats Cosmos3-Nano by a meaningful margin while running at a fraction of the compute.

Reading the RoboLab-120 Leaderboard

RoboLab-120 is a simulation-based benchmark developed by researchers at NVIDIA, the University of Toronto, and the University of Sydney, presented at RSS 2026. It contains 120 tabletop manipulation tasks built on NVIDIA Isaac Lab, organized around three competency axes: visual recognition (identifying objects by color, size, or category), procedural reasoning (stacking, reorienting, using tools), and relational logic (conjunctions, counting, spatial instructions like "to the left of"). Tasks are scored over a minimum of ten episodes each. Crucially, evaluations are run by the NVIDIA team rather than by model submitters, which prevents the gaming that briefly affected other robotics leaderboards.

The full leaderboard, as published by NVIDIA and cited by BFL, places FLUX 3 Action first among open models:

Model Type Open Success Rate Parameters FLUX 3 Action (BFL) WAM Yes 42.92% 7B OASIS WAM VLM + WAM No 39.0% — Cosmos3-Nano-Policy (NVIDIA) WAM Yes 36.8% ~16B pi0.5 (Physical Intelligence) VLA Yes 28.0% 3.3B DreamZero WAM Yes 25.7% 14B

One important clarification: the leaderboard uses an "overall" success rate across all instruction-specificity levels. A post from the RoboLab team in the months prior to this release reported Cosmos3-Nano at 39.7% on the "specific instruction" variant of the benchmark, which suggests scoring varies slightly by the granularity of the text prompt given to the model. BFL's reported 42.92% uses the same overall metric. Neither figure has been independently reproduced for FLUX 3 Action outside of BFL's own evaluation; this is expected for a model released three days before this writing.

Independent Real-Robot Results Break from Simulation

Simulation benchmarks have known limits: domain transfer from physics simulation to real hardware can be uneven, and simulation environments may not fully capture the noise and variability of physical scenes. BFL addressed this by commissioning an independent third-party evaluation from Positronic Robotics, a Romanian robotics lab, using a Franka robotic arm.

The evaluation covered ten DROID tasks, three attempts each, with a 240-second window per attempt. The operator running the robot did not know which model controlled each attempt — every run was video-recorded. Results: FLUX 3 Action completed 28 of 30 attempts (93.3%). Cosmos3-Nano completed 27 of 30 (90.0%). DreamZero completed 20 of 30 (66.7%). Pi0.5 completed 13 of 30 (43.3%).

The narrow gap between F3A and Cosmos3-Nano on real hardware (93.3% vs 90.0%) on a small task set deserves caution — 30 trials does not produce statistically robust conclusions. But the direction is consistent with the simulation results, and the evaluation's blinded design reduces experimenter bias. It also shows that F3A's simulation-trained checkpoint generalizes to a physical arm without additional real-robot fine-tuning beyond the DROID recipe.

BFL additionally fine-tuned F3A on the SO-101 arm using a small teleoperation dataset and demonstrated generalization to out-of-distribution objects, containers, and camera positions not seen during training. The SO-101 is a low-cost open-source arm, which matters because it suggests the model can anchor to new embodiments with modest data collection — a significant practical consideration for labs that cannot afford fleet-scale teleoperation.

Read more: Frontier AI models completing dangerous robot tasks instead of refusing them

Hybrid Policy Economics: When to Let the Model Think

FLUX 3 Action cannot solve every task on its own. Complex tasks requiring multi-step planning — ordering four colored blocks by a specified sequence, or finding objects partially hidden in clutter — currently exceed the capabilities of any pure action policy. GPT-6 Astra, OpenAI's embodied reasoning model, can solve all tasks in a relevant benchmark when given maximum reasoning effort, but it does so at a steep cost: approximately $13.47 per successful episode and roughly 16 minutes of wall time, according to a paper by Su et al. (2026) that BFL replicated.

BFL tested both F3A and pi0.5 as the fast-action component in a hybrid setup where the reasoning model tracks episode history and can choose to execute the action policy's predicted moves, modify a subset of them, propose its own actions, or stop. The key finding is that the quality of the action policy changes the efficiency of the entire hybrid system substantially.

With pi0.5 as the policy, the best hybrid costs $12.28 per success (medium effort Astra). With F3A as the policy, the best hybrid costs $8.77 per success (low effort Astra) and completes tasks in 8 minutes instead of 13 — 29% cheaper and 40% faster than the best pi0.5 hybrid, while still achieving 90% success across a ten-task suite that includes problems no standalone policy can solve. The interpretation is clear: a better fast policy means the reasoning model delegates more often and intervenes less, reducing token consumption and latency for every episode.

FLUX 3 Action alone, running on an H200 rented at $3 per hour, costs approximately $0.09 per successful task and takes under two minutes. Most of that time is idle, waiting for the physical robot to execute the previously predicted action chunk — meaning a single H200 instance could theoretically serve 47 simultaneous robot rollouts without additional hardware cost.

Read more: GPT-6 Astra's launch, autonomous capabilities, and AI safety monitoring

What a Non-Commercial License Means for Robotics Teams

The open-weight release comes with an important constraint. The weights are available under BFL's non-commercial license terms, which permits research and academic use but requires a separate commercial agreement for production deployment. Industrial robotics teams considering F3A as the basis for a commercial product will need to negotiate directly with BFL.

This structure mirrors how BFL has handled previous FLUX releases: open weights for the research community, commercial licensing for enterprise. It allows BFL to extract downstream value from industrial adoption while keeping the research ecosystem accessible. For academic robotics labs and non-commercial research, the weights are immediately usable via Hugging Face.

There is also a hardware consideration. The step-distilled variant — the fastest checkpoint — performs 1.15 to 2.42 times slower than pi0.5 on an RTX 5090 consumer GPU in some configurations. The guidance-distilled checkpoint is faster than Cosmos3-Nano on datacenter hardware but may not provide a consumer-GPU speed advantage. Teams building on modest hardware should test against their specific GPU and serving setup before assuming the benchmark speed figures translate directly.

From Robot Arms to Game Controllers: What BFL Is Actually Building

The most forward-looking section of BFL's technical report treats video gaming as a testbed for action prediction beyond robotics. FLUX 3 Action's midtraining includes gaming data with a 64-dimensional action space — mouse, keyboard — and BFL's write-up explicitly frames this as a step toward agents that can operate software through computer use and other latency-sensitive applications. The research parallel is clear: an agent that plays a game by predicting the next frame and the next keypress from visual context is structurally similar to an agent that operates a browser, a spreadsheet, or a development environment.

BFL's technical report acknowledges the collaboration with NVIDIA throughout development, thanking the NVIDIA Isaac Lab and RoboLab teams for the benchmarking infrastructure that underpins the model's evaluation. The two companies occupy an interesting position: NVIDIA's Cosmos3-Nano is the open model F3A now outperforms on RoboLab-120, yet both were developed in a shared ecosystem.

The physical AI benchmark wars are likely to continue. RoboLab-120 was designed with evolving task libraries specifically to prevent performance saturation — NVIDIA's team can introduce new tasks to test capabilities that current models lack. Independent researchers will begin fine-tuning FLUX 3 Action on their own robot data within weeks. Whether the video-pretraining advantage holds across more diverse real-world conditions, and whether the non-commercial license limits meaningful external testing, are the two questions that matter most for evaluating how much this benchmark result will translate to deployed robots.