Black Forest Labs Releases FLUX 3 Action, a 7B Robot Model That Tops RoboLab-120

BFL describes the model as deployable with conditions. The DROID policy needs about 32 GB of GPU memory in BF16 on an H200. It can fit 24 GB cards with FP8 quantization and text encoder offload. The FLUX Kommunity License allows non-commercial use.

FLUX 3 Action targets a trade-off in open robot policies. WAMs such as NVIDIA's Cosmos 3 Nano lead RoboLab at 36.8%, but predicting video is expensive. VLAs such as π0.5 are fast, but reach only 28.0%. On a B200, BFL measured that Cosmos 3 Nano in FP8 needs about 4.7 times more processing time than π0.5 in BF16 per second of robot motion. FLUX 3 Action keeps joint video and action prediction, and BFL closes the speed gap with a smaller backbone and distillation instead.

FLUX 3 Action is derived from the multimodal FLUX 3 backbone. Pretraining used image, video and audio data, and video made up more than 95% of training tokens. Text, video and robot state are encoded into tokens. The backbone's future tokens are decoded into video frames, and its action tokens into robot actions. Midtraining mixed pretraining data, 36.95% of samples, with action-aligned video, 63.05%. The action data covered game recordings, egocentric human hand video, handheld grippers and teleoperation across 14 embodiments. Most robot data uses a shared 50-dimension end-effector action space called EE50.

Pretraining had a large effect. Without it, DROID-only training stayed below 1% on RoboLab. With pretraining, the same protocol reached 11.6%.

RoboLab-120 has 120 tabletop tasks in Isaac Sim, with 10 trials each on a DROID-style Franka setup. FLUX 3 Action, a 7B WAM, scored 42.92%. Cosmos3-Nano-Policy, a 16B WAM, scored 36.8%. π0.5, a 3.3B VLA, scored 28.0%. DreamZero, a 14B WAM, scored 25.7%. GR00T N1.6, a 3B VLA, scored 7.2%. That is a 6.1 percentage point lead with 56% fewer parameters than Cosmos 3 Nano. The 42.92% figure is the leaderboard entry. BFL's multi-seed mean for the guidance-distilled FP8 checkpoint is 42.24% ± 0.36.

Real hardware results follow the same pattern. Positronic Robotics ran a blind evaluation on a Franka arm with 10 DROID tasks and 3 attempts each. FLUX 3 Action completed 28 of 30 attempts, or 93.3%. Cosmos 3 Nano scored 27/30, DreamZero 20/30 and π0.5 13/30.

BFL ships the DROID policy in three recipes, each in BF16 and FP8. Base uses 4 sampling steps with split guidance, video CFG 4 and action CFG 1. Guidance-distilled drops the second guidance pass, runs 1.8x to 2x faster and scores 0.6 to 1.08 percentage points higher. Step-distilled uses 1 sampling step, runs 3.15x to 4x faster, with a 3.51 to 4.32 percentage point drop in success.

Against Cosmos 3 Nano in FP8, the base and guidance-distilled checkpoints run 1.52x to 3.95x faster across consumer, workstation and datacenter GPUs. Each call yields 32 actions at 15 Hz, or 2.13 seconds of motion. π0.5 yields 1.0 second per call. BFL reports speed as real-time factor, not per-call latency. In FP8, the step-distilled checkpoint beats π0.5 by 1.34x to 2.28x on workstation and datacenter GPUs. On an RTX 5090, it is slower than π0.5.

BFL also tested hybrid control with GPT 6 Astra, following the Su et al. (2026) setup. The reasoner can execute, edit or replace the policy's predicted actions. With FLUX 3 Action and low reasoning effort, the hybrid solved 90% of episodes at $8.77 and 8 minutes 8 seconds per success. Pure Astra at maximum effort solved 100%, but cost $13.47 and 16 minutes 23 seconds per success.

AI teams can fine-tune FLUX 3 Action on their own demonstrations. BFL published a DROID recipe and an SO-101 LoRA recipe. Its documentation shows an SO-101 pick-and-place skill learned from about 200 demonstrations. With NVIDIA, BFL integrated the model natively into Hugging Face LeRobot, and it supports edge deployment on NVIDIA Jetson. The documentation also includes game-playing and simulated drone examples.

The model outputs joint targets with no built-in velocity, force or workspace limits. Applications must enforce those limits.

Editor's Summary Black Forest Labs released FLUX 3 Action, a 7B open-weights robot control model that leads the RoboLab-120 benchmark at 42.92% and outperforms larger WAMs and faster VLAs in several evaluations. It offers fine-tuning recipes and LeRobot and Jetson support under a non-commercial license, but deployment requires significant GPU memory or FP8 quantization and external safety limits.