Black Forest Labs releases FLUX 3 Action, a 7B open-weight world action model for robot control

Black Forest Labs has released FLUX 3 Action, a 7-billion-parameter open-weight world action model (WAM) derived from its multimodal FLUX 3 backbone. The model jointly predicts future video frames and robot actions, targeting both physical robotics and digital environments such as gaming. On the RoboLab-120 benchmark, the single-step checkpoint achieves a 38.3% success rate — ahead of NVIDIA's Cosmos 3 Nano at 36.8% — while running 1.34× to 2.28× faster per second of robot motion than Physical Intelligence's π0.5 on workstation and datacenter GPUs. A guidance-distilled variant pushes the success rate to 42.2% and a step-distilled checkpoint reaches 37.92% with a 3.15×–4× speedup over the base model.

The company says the efficiency gains come from two main changes: a backbone under half Cosmos 3 Nano's size, made possible by multimodal Self-Flow pretraining, and distillation that removes the guidance pass and reduces sampling to a single step without separating video from action prediction. On a B200 GPU at FP8 precision, the guidance-distilled model reaches 42.24% success at a real-time factor of 0.048, compared with 36.8% at 0.150 for Cosmos 3 Nano and 28% at 0.032 for π0.5 (served in BF16). The step-distilled variant reaches 37.92% at a real-time factor of 0.015.

Beyond raw speed, Black Forest Labs explored hybrid policies that pair FLUX 3 Action with a reasoning model — GPT 6 Astra — to handle tasks that pure action policies cannot solve alone. In the Su et al. (2026) benchmark, a hybrid using FLUX 3 Action at low reasoning effort solves 90% of episodes while cutting cost per success to $8.77 and time to 8 minutes, versus $13.47 and 16 minutes for pure Astra at maximum effort. The company says the better the fast policy, the less the system has to reason, and expects further improvements in using "as much reasoning as necessary and as little as possible."

Independent third-party evaluation by Positronic Robotics on a Franka arm across ten DROID tasks showed FLUX 3 Action completing 28 of 30 attempts (93.3%), compared with 90% for Cosmos 3 Nano, 66.7% for DreamZero, and 43.3% for π0.5. The same finetuning recipe was also applied to the SO-101 embodiment, demonstrating generalization to out-of-distribution object, container, and camera changes.

Video made up more than 95% of the training tokens in a pretraining mixture that also included images and audio. An action-focused midtraining phase incorporated gaming data (19.55% of samples), egocentric human-hand videos (13.54%), handheld gripper data (14.03%), and teleoperation across 14 embodiments (15.93%). Finetuning on DROID used a frozen-weights warmup strategy and joint finetuning with a batch size of 2,048, reaching a peak RoboLab success rate of 40.13% at 20K steps before plateauing.

Inference optimizations include row-wise dynamic FP8 quantization (1.4× speedup on local GPUs with minimal success-rate impact), prompt-token caching, and torch.compile with reduced overhead. The company says the open-weight release aims to let more researchers explore action prediction in games and simulators as a testbed for navigation, computer-use agents, and planning under partial information.

Table 1: RoboLab-120 overall success rates. Types and weight availability follow the RoboLab leaderboard

Confirmed

• FLUX 3 Action is a 7B open-weight world action model from Black Forest Labs, released September 2026.
• Single-step checkpoint: 38.3% ± 0.38 RoboLab-120 success rate; guidance-distilled: 42.2% ± 0.36; step-distilled: 37.92% ± 0.31.
• Outperforms Cosmos 3 Nano (36.8%, 16B) and π0.5 (28.0%, 3.3B) on RoboLab-120 leaderboard.
• Speed: 1.34×–2.28× faster than π0.5 (BF16) on workstation/datacenter GPUs; 1.52×–3.95× faster than Cosmos 3 Nano (FP8) across consumer, workstation, and datacenter GPUs.
• Real-time factor on B200 FP8: guidance-distilled 0.048, step-distilled 0.015 vs. Cosmos 3 Nano 0.150 and π0.5 0.032 (BF16).
• Predicts 2.13 s of motion per chunk at 5 Hz vs. π0.5's 1.0 s.
• Hybrid with GPT 6 Astra (low effort): 90% success, $8.77/success, 8 min/success vs. pure Astra $13.47, 16 min.
• Real-robot eval (Positronic Robotics, Franka arm): 93.3% (28/30) vs. Cosmos 3 Nano 90%, DreamZero 66.7%, π0.5 43.3%.
• SO-101 finetuning demonstrated with OOD object, container, and camera changes.
• Pretraining: >95% video tokens; midtraining mixture includes gaming, egocentric, handheld, teleop data.
• DROID joint finetuning (BS 2048) peaks at 40.13% at 20K steps; online Power EMA (σrel=0.10) selected.
• Inference: FP8 quantization, prompt caching, torch.compile; CFG video=4.0, action=1.0.
• Weights and report published at bfl.ai/models/flux-3-action.

Unknown

• Exact licensing terms for the open weights (not stated in the report) — not stated in the report.
• Hardware requirements for local deployment beyond tested GPUs (B200, H200, RTX 6000 Pro, RTX 5090).
• Independent replication of RoboLab and real-robot results by third parties beyond Positronic Robotics.
• Availability of training code, data pipelines, and midtraining checkpoints for reproduction.
• Long-term maintenance, update cadence, and community support commitments.

Our take

FLUX 3 Action moves the Pareto frontier for open world action models by delivering higher success rates at lower latency than both Cosmos 3 Nano and π0.5, while keeping joint video-action prediction intact. The hybrid results with GPT 6 Astra are especially notable: a 7B fast policy cuts reasoning costs by 35% and time by 50% while retaining 90% task coverage. That shifts the economic calculus for embodied AI — reasoning becomes a sporadic corrective layer rather than the primary driver. The open-weight release and detailed recipe (timestep distributions, CFG splits, EMA settings) set a new transparency bar for the field. What remains unproven is whether the single-step distillation generalizes beyond the RoboLab task distribution and DROID finetuning regime, and whether the licensing terms will allow commercial deployment without friction.

Sources

• Black Forest Labs: FLUX 3 Action — A 7B World Action Model for Robot Control