What hardware does FLUX 3 x mimic run on?

FLUX 3 x mimic runs the FLUX 3 backbone at under 80 ms on a single NVIDIA RTX 5090 GPU sitting on the robot itself, giving a full sensor-to-actuator reaction time of 101 ms. No datacenter inference is needed for the demonstrations Black Forest Labs and mimic showed at Audi facilities.

How does FLUX 3 x mimic compare with vision-language-action models?

FLUX 3 x mimic is a video-action model, not a vision-language-action model. Instead of describing the task in text first, it predicts the next frames of the scene and reads joint commands out of that latent stream. Mimic reports up to 10× better sample efficiency than typical vision-language-action baselines on soft-body factory tasks.

Are the FLUX 3 x mimic weights open?

The base FLUX 3 backbone is available with open weights on Hugging Face and via API from Black Forest Labs, but the mimic-specific action decoder and reinforcement-learning post-training rely on mimic robotics' internal teleoperation data. The blog announcement points to mimic's product page rather than a public model card for the joint version.

Where has FLUX 3 x mimic been deployed?

Mimic robotics tested FLUX 3 x mimic on soft-body kitting tasks inside Audi production facilities, using its teleoperated robots and the U1 wearable data-collection system. The published numbers — 95% success rate, 101 ms reaction time, 3,500 training steps to recover full video-generation quality — come from those factory pilots.

Black Forest Labs · 2026-07-23 · major

FLUX 3 x mimic — the FLUX 3 backbone drives factory robots at 101 ms

Black Forest Labs and mimic robotics released FLUX 3 x mimic, a video-action model that runs on a single RTX 5090 and hits 101 ms end-to-end reaction time in Audi factory tests, with 95% success on soft-body kitting.

FLUX 3 x mimic banner from Black Forest Labs' announcement blog

A FLUX 3 spin-off that lets one on-robot GPU pilot a real factory arm in near real time.

Quick facts

Makers	Black Forest Labs + mimic robotics
Backbone	FLUX 3 video model
On-robot latency	<80 ms on a single NVIDIA RTX 5090
End-to-end reaction time	101 ms
Success rate	95% on soft-body kitting tasks
Sample efficiency	Up to 10× vs. VLA baselines
Field test	Audi production line pilots

What is it?

FLUX 3 x mimic is a video-action model from Black Forest Labs and mimic robotics that decides what a robot should do next by generating what it should see next. The joint model reuses the FLUX 3 video backbone and adds a custom action decoder, so one network handles both perception and control.

How does it work?

The action decoder attends to the FLUX 3 backbone's latent video stream and outputs joint commands in overlapping chunks, so the next prediction runs while the current one is still executing. Mimic post-trains with reinforcement learning on data from its teleoperated robots and U1 wearable, then quantizes the model to run inference in under 80 ms on a single NVIDIA RTX 5090.

Why does it matter?

Robotics teams have needed datacenter GPUs or heavy sim-to-real pipelines to push a single manipulation demo past 90% success. FLUX 3 x mimic clears 95% on soft-body kitting tasks with roughly 10× fewer demonstrations than typical vision-language-action models, and runs cheaply enough to sit on the robot itself — closer to plug-and-play deployment on a factory line.

Who is it for?

industrial robotics teams, embodied-AI researchers

Frequently asked questions

What hardware does FLUX 3 x mimic run on?: FLUX 3 x mimic runs the FLUX 3 backbone at under 80 ms on a single NVIDIA RTX 5090 GPU sitting on the robot itself, giving a full sensor-to-actuator reaction time of 101 ms. No datacenter inference is needed for the demonstrations Black Forest Labs and mimic showed at Audi facilities.
How does FLUX 3 x mimic compare with vision-language-action models?: FLUX 3 x mimic is a video-action model, not a vision-language-action model. Instead of describing the task in text first, it predicts the next frames of the scene and reads joint commands out of that latent stream. Mimic reports up to 10× better sample efficiency than typical vision-language-action baselines on soft-body factory tasks.
Are the FLUX 3 x mimic weights open?: The base FLUX 3 backbone is available with open weights on Hugging Face and via API from Black Forest Labs, but the mimic-specific action decoder and reinforcement-learning post-training rely on mimic robotics' internal teleoperation data. The blog announcement points to mimic's product page rather than a public model card for the joint version.
Where has FLUX 3 x mimic been deployed?: Mimic robotics tested FLUX 3 x mimic on soft-body kitting tasks inside Audi production facilities, using its teleoperated robots and the U1 wearable data-collection system. The published numbers — 95% success rate, 101 ms reaction time, 3,500 training steps to recover full video-generation quality — come from those factory pilots.