FLUX 3 Action

FLUX 3 Action is an open weights 7B world action model from Black Forest Labs. It takes camera frames, the robot's state and a text instruction, and returns the next chunk of actions, denoised together with the next video frames.

This repository is the standalone training and inference code for robots, simulators and games.

Weights · Documentation · Blog post · LeRobot integration

transformer.py, transformer_inf_bf16.py and transformer_inf_fp8r.py keep training and inference implementations separate so readers, especially coding agents, can follow each path directly without tracing inheritance or configuration branches.

The FLUX 3 Action collection contains separate DROID and SO-101 policy repositories. Their frozen encoders are shared through the base repository.

DROID offers BF16 and FP8r packages for base, guidance-distilled and single-step inference.

Recipe	BF16 subfolder	FP8r subfolder
Base: 4 steps with guidance	root (omit subfolder)	variants/fp8r
Guidance-distilled: 4 steps	variants/gd	variants/gd-fp8r
Step-distilled: 1 step	variants/sd	variants/sd-fp8r

Standalone FLUX Action supports all six packages. BF16 packages also include LeRobot configs; native FP8r packages require FLUX Action. LeRobot reads config.json; FLUX Action reads config.native.json when present, otherwise config.json. See checkpoint selection for commands.

Get started

Guide	What you’ll do
Setup & inference	Install dependencies, download weights and predict actions with DROID or SO-101
DROID full fine-tuning	Prepare DROID data, run the reproduction recipe, resume and export
SO-101 LoRA	Adapt the prepared SO-101 checkpoint to your task with LeRobot
Custom tasks and embodiments	Define your own observations and actions, using games as a worked example