Model card

Made by
Black Forest Labs
Size
7B parameters
Published
22 Sep 2026
More about the model
Publication
Technical report
Weights licence
FLUX Kommunity License v1.0, non-commercial source
Actions per chunk
32, at 15 Hz source
Smallest GPU
RTX 5090, the smallest GPU in Black Forest Labs' timing table source
Inference time
136 ms per chunk on a B200 (BF16, guidance-distilled) source
Architecture
Diffusion transformer that denoises video frames and actions together, with a frozen video VAE and a Qwen3-VL-4B text encoder source
Training data
Video, image and audio pretraining; action training on games, human-hand video, handheld grippers and teleoperation from 14 embodiments; fine-tuned on DROID source

Run with Positronic

Start the server source

docker run --gpus all -p 8000:8000 positro/flux3-action

Drive a DROID robot from a positronic checkout guide

uv run --extra hardware positronic-inference real \
  --policy=.remote \
  --policy.address.host=localhost --policy.address.port=8000 \
  --next_task.instruction="Put the marker on the plate"

Progress through an episode

Each band shows how many episodes of FLUX 3 Action had succeeded or had no success at each second of simulated time. An episode that has ended stays in its band. The dashed line shows the mean Reliability, the share of the episodes that had succeeded.

Episodes

Each episode of FLUX 3 Action on MolmoSpaces-320, the newest first.