BFL

FLUX 3 Action

Black Forest Labs, Germany 路 September 2026

World model 路 Open weights, FLUX Kommunity License 路 Also written F3A, FLUX.3 Action, flux-3-action-base, flux-3-action-droid, flux-3-action-so101

Parameters
7 B
7B according to the announcement and the RoboLab leaderboard, less than half the size of Cosmos 3 Nano at 16B.
GPU memory
69 GB
reported by the authors
Inference latency
not published
per action step
Weights
FLUX Kommunity License
Weights under the FLUX Kommunity License, gated on Hugging Face with automatic approval after accepting the terms; read the...

What it is

FLUX 3 Action is Black Forest Labs' open-weight world action model: a 7B model derived from the multimodal FLUX 3 backbone that predicts future video and robot actions jointly. Its DROID policy returns 32 joint position actions per call, 2.13 s of motion, and distillation removes the guidance pass and cuts sampling to a single step without separating video from action. On NVIDIA's RoboLab-120 it scores 42.9 percent, against 36.8 percent for the 16B Cosmos3-Nano-Policy and 28.0 percent for 蟺0.5, and in a blind third-party test on a real Franka it completed 28 of 30 attempts against 13 of 30 for 蟺0.5. In a later test by Runway on the same tasks it reached 80 of 100 episodes against 88 for Praxis-1. The weights are gated on Hugging Face under the FLUX Kommunity License, the code is Apache-2.0, and there is a prepared SO-101 checkpoint with a LoRA recipe for LeRobot.

Architecture

Backbone
FLUX 3, Black Forest Labs' multimodal generative backbone, pretrained on image, video and audio data with video accounting for over 95 percent of the training tokens. No vision language model backbone.
Action head
Joint video and action prediction in one model, without a separate action decoder. Three recipes are released, each in BF16 and FP8: base with 4 steps and guidance, guidance-distilled with 4 steps, and step-distilled with 1 step. The DROID policy returns 32 actions per call, each with seven absolute joint targets and a gripper value, covering 2.13 s of motion. The SO-101 policy predicts 42 actions and executes the first 32 at 30 Hz before it replans.
Action chunk
32 steps per forward pass
Control rate
15 Hz
Parameters
7B according to the announcement and the RoboLab leaderboard, less than half the size of Cosmos 3 Nano at 16B. Chunk length and rate above are for the DROID policy: 32 actions for 2.13 s of motion correspond to 15 Hz.
Pretraining data
FLUX 3 pretraining on image, video and audio, over 95 percent video by tokens. Action mid-training mixture by samples: 36.95 percent joint video and audio, 19.55 percent gaming, 13.54 percent egocentric human hands, 14.03 percent handheld and 15.93 percent teleoperation data. The DROID policy is the base model fine-tuned on DROID demonstrations.
Embodiments
Franka (DROID setup), SO-101 arm

What hardware it needs

This is the section most people came for, so it is worth being precise about which numbers are measured and which are arithmetic.

Computed from 7 B parameters. Weights only, so treat it as a floor: activations, image encoder intermediates and the CUDA context come on top, in practice 30 to 50 percent more for inference.
PrecisionWeightsNote
fp3226 GBTraining master weights, rarely used for inference
bf16 or fp1613 GBThe usual way these checkpoints are served
int86.5 GBQuantised, expect some loss of precision on fine motions
int43.3 GBAggressive quantisation, verify success rate before trusting it

Smallest real card that fits the bf16 weights plus a 40 percent runtime allowance: 24 GB. That is an estimate for inference, not for fine tuning. Full fine tuning also holds optimiser state and gradients, which typically costs several times the weights unless you use LoRA or a comparable adapter.

Published figures. These come from the model authors, not from this site.
Inference memory, reported69 GB source
Control rate15 Hz

The RoboLab leaderboard lists 69 GiB for the evaluated DROID policy. The setup guide names an NVIDIA H200 with CUDA 12.8; inference and training need Linux and an NVIDIA GPU. FP8 packages exist for the DROID configuration. No latency per call is published, only real-time factors: on a B200 in FP8 the guidance-distilled checkpoint runs at 0.048 and the step-distilled one at 0.015, against 0.150 for Cosmos 3 Nano and 0.032 for 蟺0.5 in BF16.

Reported results

Grouped by what the evaluation actually asked. Values from different suites are not comparable with each other, so each one keeps its suite and split.

Simulation

Reproducible benchmark suites. Everyone runs the same episodes, so the numbers can be compared inside a suite.

  • RoboLab-120 overall, 10 episodes per task
    42.9%
    success rate515 of 1,200 episodes, marked as verified by the RoboLab organizers. Cosmos3-Nano-Policy 36.8, 蟺0.5 28.0 percent on the same board.source
  • RoboLab-120 step-distilled checkpoint, single sampling step
    38.3%
    success rateAuthors' own run, plus or minus 0.38. The guidance-distilled checkpoint reaches 42.2 percent.source

Real world

Rollouts on physical hardware. The setups differ, so read these as evidence, not as a ranking.

  • Real robot, Franka at Positronic (ten single-item tasks) 3 attempts per task, blind third-party evaluation
    93.3%
    success rate28 of 30 attempts, run by Positronic Robotics; the operator did not know which model drove each attempt. Cosmos 3 Nano 27 of 30, DreamZero 20 of 30, 蟺0.5 13 of 30 under the same protocol.source
  • Real robot, Franka at Positronic (ten single-item tasks) 10 episodes per task, Runway's test of November 2, 2026
    80%
    success rateMeasured by Runway, not by the developer: 80 of 100 episodes on a Franka FR3. Praxis-1 88 of 100 on the same rig.source

Fine tuned tasks

No results in this category are published for this model.

On real hardware

Positronic Robotics ran the policy blind on a Franka arm: 28 of 30 attempts on ten DROID tasks, against 27 for Cosmos 3 Nano, 20 for DreamZero and 13 for 蟺0.5. In Runway's later test of the same tasks it reached 80 of 100 episodes, Praxis-1 88. In the real-world trials of Apple's AMP1 paper it succeeded in 49 of 108. Black Forest Labs also shows SO-101 rollouts after fine-tuning on a small teleoperated dataset, including changed objects, containers and camera positions, without success rates.

Fine tuning it yourself

The repository has a full fine-tuning recipe for DROID and a LoRA recipe for the SO-101 through LeRobot. The SO-101 recipe expects a 30 Hz LeRobot v2.1 or v3.0 dataset with a fixed scene camera and a wrist camera, and its preset uses rank and alpha 32, one GPU with batch 2 and accumulation 4, and 10,000 microbatches (2,500 optimizer updates). No demonstration dataset is included, you bring your own. Custom embodiments and action spaces are documented separately.

Where it helps, where it does not

Strengths

  • 42.9 percent on RoboLab-120 with 7B parameters, ahead of the 16B Cosmos3-Nano-Policy at 36.8 and 蟺0.5 at 28.0 percent.
  • Efficient for a world action model: the step-distilled checkpoint needs a single sampling step and still reaches 38.3 percent.
  • Blind third-party test on a real Franka: 28 of 30 attempts, against 13 of 30 for 蟺0.5.
  • Prepared SO-101 checkpoint and LoRA recipe with LeRobot integration; the code is Apache-2.0.

Limits

  • 69 GiB of GPU memory for the evaluated DROID policy, according to the RoboLab leaderboard.
  • The weights are gated and under the FLUX Kommunity License, not a standard open source licence.
  • Later head-to-head results are lower: 80 of 100 against 88 for Praxis-1 on the same real tasks, and 49 of 108 in the trials of Apple's AMP1 paper.
  • Fine-tuned by Runway on RT-1 data it reached 39.7 percent on SimplerEnv, against 80.0 percent for Praxis-1; that number is a competitor's measurement.
  • No success rates are published for the SO-101 checkpoint.

Sources

Everything on this page was taken from these documents. Where they disagree with what you read here, they win.

Entry last checked 2026-11-22.

Try a policy on a real arm

A real SO-100 runs on /live and can be driven from the browser when it is online. Five of the models in this arena can be fine tuned on a dataset you record with your own arm.