Robotics Breakthrough: Black Forest Labs' FLUX 3 Action, a 7B AI, Dominates Control Benchmarks

Black Forest Labs (BFL), known for its FLUX image generation models, has officially released FLUX 3 Action. This new offering represents a significant advancement in robot control, delivered as a 7-billion-parameter, open-weights World Action Model (WAM). The model functions by interpreting real-time camera feeds, current robot states, and textual instructions, subsequently predicting upcoming video frames and corresponding sequences of robot actions. Demonstrating its prowess, FLUX 3 Action achieved a leading 42.92% task success rate on the rigorous RoboLab-120 benchmark.

Deployment Considerations

The practical deployment of FLUX 3 Action, specifically its DROID policy, is feasible under certain conditions. Optimal performance in BF16 precision typically necessitates around 32 GB of GPU memory on high-end hardware like an H200. However, the model can also be adapted for GPUs with 24 GB of memory by employing FP8 quantization and offloading the text encoder. The FLUX Kommunity License governs its use, permitting non-commercial applications.

Bridging the Performance-Efficiency Gap

Traditionally, open-source robot policies have presented a dilemma between performance and efficiency. World Action Models, such as NVIDIA’s Cosmos 3 Nano, often achieve superior results, leading RoboLab with scores around 36.8%, but their video prediction capabilities incur substantial computational costs. Conversely, Vision-Language-Action (VLA) models like π0.5 prioritize speed, yet typically exhibit lower success rates, approximately 28.0%. For instance, BFL's internal metrics indicated that Cosmos 3 Nano (FP8) required roughly 4.7 times more processing time per second of robot movement compared to π0.5 (BF16) on a B200 processor. FLUX 3 Action aims to reconcile this trade-off by retaining integrated video and action prediction while minimizing the speed deficit through a more compact backbone architecture and distillation techniques.

Architecture and Training Insights

The foundation of FLUX 3 Action is the multimodal FLUX 3 backbone. Its pretraining involved diverse data streams, including images, video, and audio, with video data constituting over 95% of the training tokens. During operation, text, video, and robot state information are converted into tokens. The backbone then processes these tokens, with future tokens being translated back into predicted video frames and action tokens into precise robot movements. Mid-training protocols blended pretraining data (36.95% of samples) with action-aligned video (63.05%). This action dataset encompassed a wide range of scenarios, including game recordings, human egocentric hand footage, handheld gripper operations, and teleoperation across 14 distinct robot embodiments. A shared 50-dimension end-effector action space, referred to as EE50, was utilized for most robot data. The importance of pretraining is stark, as DROID-only training without this initial phase yielded less than 1% success on RoboLab, whereas the identical protocol achieved 11.6% with pretraining.

Benchmark Dominance

The RoboLab-120 benchmark comprises 120 distinct tabletop tasks simulated within Isaac Sim, with each task undergoing 10 trials using a DROID-style Franka setup. FLUX 3 Action recorded a leading 42.92% success rate, surpassing Cosmos 3 Nano (WAM, 16B parameters) at 36.8% and π0.5 (VLA, 3.3B parameters) at 28.0%. This represents a notable 6.1 percentage point advantage over Cosmos 3 Nano, despite FLUX 3 Action utilizing 56% fewer parameters. While the leaderboard entry shows 42.92%, BFL's multi-seed average for the guidance-distilled FP8 checkpoint stood at 42.24% ± 0.36. Real-world evaluations conducted blindly by Positronic Robotics on a Franka arm across 10 DROID tasks (3 attempts each) corroborated these findings. FLUX 3 Action successfully completed 28 out of 30 attempts (93.3%), outperforming Cosmos 3 Nano (27/30), DreamZero (20/30), and π0.5 (13/30).

Optimized for Speed: Three Checkpoints

BFL offers the DROID policy in three distinct configurations, each available in BF16 and FP8 precision:

• Base: Utilizes 4 sampling steps with split guidance (video CFG 4, action CFG 1).
• Guidance-distilled: Omits the second guidance pass, resulting in 1.8x to 2x faster operation and a 0.6 to 1.08 percentage point improvement in success rate.
• Step-distilled: Employs a single sampling step, achieving 3.15x to 4x faster execution, though with a 3.51 to 4.32 percentage point reduction in success.

Compared to Cosmos 3 Nano in FP8, the base and guidance-distilled checkpoints of FLUX 3 Action demonstrate 1.52x to 3.95x faster performance across various GPU categories, including consumer, workstation, and datacenter. Each model invocation generates 32 actions over 2.13 seconds of motion at 15 Hz, a longer planning horizon than π0.5's 1.0 second per call. When evaluating real-time factor, the step-distilled FP8 checkpoint surpasses π0.5 by 1.34x to 2.28x on workstation and datacenter GPUs, though it performs slower on an RTX 5090.

Hybrid Control with AI Reasoners

Further exploration by BFL included hybrid control experiments with GPT 6 Astra, mirroring the setup described by Su et al. (2026). In this configuration, the reasoner component possesses the ability to execute, modify, or completely supersede the policy's predicted actions. With FLUX 3 Action and minimal reasoning overhead, the hybrid system successfully resolved 90% of episodes, averaging $8.77 and 8 minutes 8 seconds per success. For comparison, a pure Astra approach with maximum reasoning effort achieved 100% success but incurred a higher cost of $13.47 and a longer average time of 16 minutes 23 seconds per successful episode.

Fine-Tuning, LeRobot, and Edge Integration

AI development teams can fine-tune FLUX 3 Action using their own custom demonstration datasets. BFL has provided a DROID recipe and an SO-101 LoRA recipe, with documentation illustrating an SO-101 pick-and-place skill acquired from approximately 200 demonstrations. In collaboration with NVIDIA, the model has been seamlessly integrated into Hugging Face LeRobot and supports deployment on NVIDIA Jetson edge devices. The accompanying documentation also features examples related to game playing and simulated drone control. It is important to note that the model outputs joint targets without integrated velocity, force, or workspace limits; thus, enforcing these constraints falls to the implementing application.

Key Takeaways

• FLUX 3 Action, a 7B World Action Model, leads the RoboLab-120 benchmark with a 42.92% success rate.
• The model achieves a 6.1 percentage point lead over Cosmos 3 Nano while using 56% fewer parameters.
• It operates up to 3.95 times faster than Cosmos 3 Nano in FP8 precision.
• FLUX 3 Action plans 2.13 seconds of motion per call, offering a longer horizon than π0.5's 1.0 second.
• Its weights, code, and recipes are available under the FLUX Kommunity License.