Flux 3 Action

The image-model lab shipped a robot brain. Flux 3 Action is a 7B open-weight world-action model derived from Black Forest Labs’ FLUX 3 multimodal backbone: camera frames, robot state, and a text instruction go in; the next chunk of actions comes out denoised together with predicted video frames. It was pretrained on video, image, and audio at scale (video-heaviest), then adapted to a target robot through joint video-action training. On the RoboLab-120 leaderboard the guidance-distilled checkpoint sets a state-of-the-art success rate of 42.2% (+/- 0.36) with less than half the parameters of the previous best open model, running up to 3.95x faster than Cosmos 3 Nano in FP8.

What runs where. The weights are 16.5 GB total (13.97 GB base + 2.51 GB video VAE) in bf16, so a 24 GB RTX 4090 or RTX 6000 Pro holds the base model; FP8 quantization is where the vendor does its speed runs. The latency table on B200: 136ms per 4-step chunk with guidance off (102ms FP8), 41ms at 1 step (32ms FP8). The honest caveat: on workstation and datacenter GPUs F3A is 1.34-2.28x faster than pi0.5, but on consumer cards like the RTX 5090 it is 1.15-2.42x SLOWER than pi0.5 - the consumer-GPU story is real but not the headline. Output format matters for control loops: 32 actions per prediction at 15 Hz with a 2.13-second horizon, versus pi0.5’s 15 actions at 1.0 second.

License, read carefully. Flux Kommunity License: free for non-commercial use for everyone (including non-commercial robotics); commercial use - including any Robotics Use in production - only for Qualifying Users, defined as under $5M gross annualized revenue including affiliates. A robotics startup under that line can ship product with it; anything larger needs a negotiated license. The embodiment-specific fine-tunes (so101, droid) and the training recipe are published, which matters more in robotics than in text: the recipe is the hard part.

The hybrid finding. F3A paired with a frontier reasoner (their Astra experiments) beats pure action policies by 2x and 1.54x on the hardest task class while retaining about 90% success - fast local control plus slow cloud reasoning is the architecture this benchmark is pushing toward. For robot builders: the weights and recipe are downloadable today; the license is the constraint to check before shipping.