Black Forest Labs releases FLUX 3 Action robot model

Black Forest Labs announced FLUX 3 Action (F3A), a 7B-parameter "world action model" (WAM) that predicts robot actions and future video together. The company is publishing the weights with a technical report that covers pretraining, midtraining, fine-tuning and inference optimisation.

The company is pitching F3A as a way out of a familiar trade-off. WAMs such as Nvidia's Cosmos 3 Nano lead the RoboLab benchmark (36.8% task success) but are slow. On a B200, Cosmos 3 Nano in FP8 needs about 4.7x the processing time per second of robot motion that Pi0.5, the strongest open vision-language-action model (28.0%), needs in BF16.

BFL says its single-step checkpoint reaches 38.3% on RoboLab and is 1.34x to 2.28x faster than Pi0.5 per second of robot motion on workstation and datacenter GPUs. A guidance-distilled checkpoint reaches 42.2%. The company says the speed gain is in real-time factor, not per-call latency, because F3A predicts 2.13 seconds of motion per call against Pi0.5's 1.0 seconds. BFL credits two factors: a backbone less than half the size of Cosmos 3 Nano, built with its multimodal Self-Flow pretraining, and distillation that removes the guidance pass.

The report also prices out hybrid systems in which a reasoning model hands control to a fast policy. GPT 6 Astra alone solves every task in one benchmark but costs $13.47 and about 16 minutes per success. F3A alone costs $0.09 per success on a $3/hour H200 but cannot solve every task. BFL says pairing the two still solves 90% of episodes and gives up to 53.64% more success per dollar. The provided excerpt cuts off before the full hybrid figures.

BFL also mentions gaming as a testbed for navigation and future computer-use agents, and a blind third-party evaluation on real robots. The excerpt gives no details or results for either.