# FLUX.3 Mimic
A video-action model for robotic manipulation, built by Black Forest Labs with mimic robotics on top of the [[FLUX.3]] backbone. Announced July 2026, already running on production lines at Audi.
This is the variant I flagged as the strange item in the [[FLUX.3]] note. It deserves its own note, because the idea behind it is bigger than the product.
## The core idea: actions are just another modality
BFL's argument, and it's a good one:
Video prediction consumed **over 95% of [[FLUX.3]]'s training cost**. To predict the next frame well, a model has no choice but to learn contact, motion, weight, friction, and cause and effect. You cannot fake plausible video of a hand picking up a cable without having internalized something about how cables behave.
That internalized something is a world model. It was never the goal, it was a side effect of the compression problem.
So once [[FLUX.3]] had it, adding robot control turned out to be cheap. They **train a lightweight action decoder on top of intermediate features extracted from the video prediction path**. No new foundation model. A small head on an existing one, plus a brief adaptation, after which full video generation quality returns.
The framing they use: video and audio already encode the underlying physical reality, so actions are simply another modality describing the same thing.
## Training scale
- Tens of millions of hours of general video, for broad world dynamics
- Hundreds of thousands of hours of manipulation-focused video, for the specific domain
The method is Self-Flow, the same approach unifying generation and representation learning across the [[FLUX.3]] family.
## What it does
- **Sample efficiency**: hits target success rates in roughly half the training steps of earlier approaches
- **Failure recovery**: robots correct their own mistakes without being explicitly taught to
- **Soft materials**: handles seals and cables, which are the classic hard case for manipulation
- **Task range**: kitting, insertion assembly, component assembly, flexible materials
- **Deployed**: validated on Audi production lines, not just in a lab
The failure recovery point is the one to sit with. Nobody demonstrated recovery behaviour. It emerged from a model that understands what should happen next, noticing that it didn't.
## Latency
Input to world representation in **under 80ms** on an NVIDIA RTX 5090, giving a full system reaction time around **101ms**. Roughly human visual reaction speed, on one consumer-class GPU.
## Reception
The [[Hacker News]] thread was informed and appropriately unimpressed by the novelty claim:
- The general recognition that video models contain extractable world representations is correct, and this is well-trodden ground. Nvidia, Waymo, Luma Labs and Runway have all demonstrated versions of it
- Prior art cited: Google's VLA-based robot replacing a timing belt over a year ago, and Generalist AI's dexterous manipulation work
- The **hand design** drew genuine interest, looking more sophisticated than the standard grippers on competing systems such as Xiaomi's Robotics-1
- Several commenters hoped BFL stays European-owned rather than being acquired by an American company
One comment named the business logic cleanly: train a video model, sell video generation, use the scale that funds to train robot models. Media generation pays for the world model, and the world model is the actually valuable asset.
## Why this matters beyond robotics
If actions are just another modality on top of video prediction, then the expensive part of robotics was never the robotics. It was acquiring a world model, and the generative video industry has been building those by accident while chasing something else entirely.
That reframes what a company like BFL is. It looks like an image and video generation company. The training run says it is a world model company that currently monetizes through media.
## Caveats
- No parameter counts, no technical report, no independent replication
- "Deployed at Audi" is BFL's claim about scope we cannot see. Pilot and production line are different words
- Sample efficiency and success rates are self-reported against unnamed earlier approaches
- Availability beyond the mimic partnership is unspecified
## References
- Announcement: https://bfl.ai/blog/flux-3-mimic
- [[Hacker News]] discussion: https://news.ycombinator.com/item?id=49033127
- FLUX 3 announcement: https://bfl.ai/blog/flux-3
- mimic robotics: https://www.mimicrobotics.com/
## Related
- [[FLUX.3]]
- [[FLUX.2]]
- [[AI image generation models]]
- [[Hacker News]]