Black Forest Labs
FLUX 3 Action
What FLUX 3 Action’s own documents say it is built from — every method with the sentence that describes it.
- Parameters
- 7B
- Attention
- World-action model built on a video-generation backbone
- Input
- text+image -> action
- Licence
- FLUX Kommunity License v1.0 (gated)
- Released
- 2026-09-23
Against the consensus recipe
Its documents state 0 of the 18 methods the field agrees on.
Data curation
- Benchmark decontaminationnot in its documents
- Synthetic data generationnot in its documents
- Corpus deduplicationnot in its documents
Training objective
- Multi-Token Predictionnot in its documents
- Knowledge insulationnot in its documents
- Quantization-aware distillationnot in its documents
Model architecture
- Mixture of expertsnot in its documents
- Multi-Token Predictionnot in its documents
- Grouped-query attentionnot in its documents
Optimization
- Muonnot in its documents
- Expert parallelismnot in its documents
- Quantization-Aware Trainingnot in its documents
Post-training
- Supervised fine-tuningnot in its documents
- Asynchronous reinforcement learningnot in its documents
- Multi-Teacher On-Policy Distillationnot in its documents
Inference & serving
- Thinking modenot in its documents
- Multi-Token Predictionnot in its documents
- FP8not in its documents
filed at the root 1
“The model is a 7 billion parameter diffusion transformer.”
multimodal architecture 4
“New embodiments require their own action heads.”
“Each new robot needs its own input and output layers.”
“The base has 457 tensors, with video and text streams, a shared action trunk, and ee50/gaming action heads.”
“Action base weights for embodiment adaptation, plus the shared frozen encoders used by the SO-101 and DROID policies.”
data sourcing 1
“The rest came from sources such as video game play and first-person footage of human hands”
data mixture & curriculum 1
“whose pretraining tokens were more than 95 percent video”
supervised fine-tuning 6
“a frozen copy of Qwen3-VL-4B processes instructions”
“A frozen video encoder processes camera frames”
“In a second training stage focused on actions, 15.93 percent of the samples came from teleoperated robots.”
“It also fine-tuned the model to play two simple video games and fly a simulated drone.”
“BFL adapted the model to a low-cost SO-101 arm using about 200 demonstrations.”
“To train heads for a new robot, simulator or game, follow Fine-tune.”
filed at the root 1
“Validate on a simulator or with the arm's safety limits engaged before running near people.”
3 mentioned or evaluated
human & real-world evaluation “In blind tests that BFL arranged with Positronic Robotics, a Franka arm running FLUX 3 Action completed 28 of 30 attempts”
benchmark “It ranks first on RoboLab-120, completing 42.9 percent of 1,200 trials.”
benchmark “RoboArena, which ranks models by blind head-to-head comparisons on real robot arms at academic labs.”
unfiled 7
“FLUX 3 Action is a world-action model (WAM). Given camera images and a written instruction, it predicts the robot’s next moves along with…”
“Actions form a second stream of tokens that the model denoises together with future video frames.”
“The robot carries out the first few actions, then the model looks again and replans.”
“Nothing in the model bounds joint velocity, force or workspace; the application must enforce those limits”
“the application must enforce those limits and keep a hardware stop within reach.”
“15.93 percent of the samples came from teleoperated robots.”
Not used 1
“To control a machine in a way that endangers people, without human oversight and a means to stop it.”
Sources
The curated source list for this model. "not fetched" means the URL is recorded but its text was not read in the current run.
| Source | Kind | Publisher | Status |
|---|---|---|---|
| huggingface.co/black-forest-labs/flux-3-action-base | model card | Black Forest Labs | read |
| the-decoder.com/black-forest-labs-launches-flux-3-action-an-open-roboti… | news | The Decoder | read |
| deeplearning.ai/the-batch/video-models-steer-robots | news | DeepLearning.AI | read |