FLUX 3 is running on robots inside Audi factories

Generative video models have long been viewed as tools for Hollywood VFX or social media creators. Black Forest Labs just blew that assumption wide open by sending its newest foundation model straight to an Audi assembly line.

The German AI lab introduced FLUX 3, a multimodal foundation model that generates video, images, and native synced audio. Alongside Zurich-based mimic robotics, BFL also unveiled FLUX-mimic, a specialized video-action model powered by FLUX 3 that is already handling live production tasks inside Audi factories.

Teaching physics through pixels

While FLUX 1 and 2 focused strictly on images, FLUX 3 was jointly trained on images, video, and audio from day one. Over 95% of the total compute went directly into video prediction. In comparison, audio accounted for less than 0.5% of the tokens in a 720p clip.

BFL’s core thesis is simple: to generate believable video, an AI has to master real-world physics, including weight, inertia, contact, and causality. If a model understands how objects react when pushed or pulled, predicting robotic action is simply another way of expressing that same physical reality.

When action prediction was first introduced into the training curriculum, human ratings for video generation initially dipped by up to 10%. But after just 3,500 training steps, video quality fully rebounded while retaining action capabilities. Built on BFL's Self-Flow architecture, a lightweight decoder reads features directly from the video prediction path, running on a single RTX 5090 GPU in under 80ms.

Half-hour training and real-time self-correction

At Audi, FLUX-mimic is tackling tasks that traditional factory automation historically fumbles.

Robots powered by the model are kitting components into trays, sliding electronic control units into tight-tolerance fixtures, and assembling hardware. Crucially, they can manipulate flexible, deformable materials like rubber seals and wiring harnesses, materials rigid industrial bots struggle to grasp.

Because the underlying backbone already understands physical interactions, teaching the robot a brand-new factory task now takes about 30 minutes of demonstration data instead of the conventional 30-plus hours. Furthermore, if the gripper misses a component, FLUX-mimic naturally detects the mistake and corrects its grip autonomously without halting the assembly line.

Why it matters

We think BFL and mimic have delivered one of the most compelling validations yet for physical AI: the ultimate robot brain might actually be a world simulator built for video.

For enterprise manufacturers and robotics teams, the implications are immediate. Slashing demonstration requirements from 30 hours to 30 minutes radically lowers the cost of deploying automation for high-mix, low-volume manufacturing runs. Running on a single consumer-grade workstation card also removes massive infrastructure bottlenecks from factory floors.

For the broader AI ecosystem, this threatens the dominance of traditional Vision-Language-Action (VLA) models. LLMs excel at semantic reasoning, but text models struggle with continuous spatial dynamics. Video prediction models, by contrast, spend massive compute budgets mastering spatial collisions and motion.

That said, high-stakes manufacturing leaves zero room for generative hallucinations. While autonomous recovery handles minor misses, tiny physical miscalculations on a rapid assembly line can still damage expensive equipment. For now, it looks ideal for flexible kitting and assembly, but mission-critical precision will require extensive long-term reliability testing.