Black Forest Labs released FLUX 3 on Thursday, and the headline capability is not the one the company leads with. Yes, it generates video — up to 20 seconds, with synchronised dialogue, sound effects and music produced in the same pass. But the more consequential claim is buried a layer down: the same backbone, with a small decoder attached, is currently driving robot arms at Audi on tasks the carmaker says conventional automation could not do.

The Freiburg lab is best known for images. Its founders wrote the papers behind VQGAN, latent diffusion and Stable Diffusion, and the FLUX family now sits inside Adobe Photoshop, Picsart and Canva. FLUX 1 and FLUX 2 generated pictures. FLUX 3 is the company's move out of that category entirely.

One architecture, four modalities

FLUX 3 is trained jointly across images, video and audio in a single architecture, built on an approach Black Forest Labs calls Self-Flow. The company's argument is that these are not separate problems that happen to share a codebase. In its own framing, images convey structure, video teaches dynamics, audio "reveals causal relationships between mechanical phenomena and acoustics that vision alone cannot detect" — and actions are simply one more low-dimensional signal, "a low dimensional representation of a robot's state, tightly coupled to visual observations."

The economics of that are stark in the lab's own numbers. Audio, it notes, makes up "less than 0.5% of the tokens in a 720p video with audio." Once a model has done the expensive work of learning video, the cheaper modalities come along for close to free — and, the company argues, actions follow the same shape.

Co-founder and CEO Robin Rombach put the thesis bluntly in the launch release: "You can't cheat reality. A model that only learns images can only generate images. But the world is not made of still frames. It moves, sounds, changes, and responds."

The robot is the interesting part

The robotics work is a collaboration with mimic robotics, a Zurich-based ETH spinout founded in 2024 that builds its own robots and was among the first partners given FLUX 3 access. The joint model is called FLUX-mimic, and its architecture is deliberately minimal: rather than retraining a general model for control, it trains "a lightweight action decoder on top of intermediate features extracted from the video prediction path" of FLUX 3.

That design choice produces the launch's strongest technical claim. Black Forest Labs reports that the action decoder "outperforms previous vision-language-action models, even with a completely frozen FLUX backbone — a setting where previous vision-language-action models fail to succeed." In plain terms: the video model's internal representation of how the world behaves is, on the company's benchmarks, already good enough to control a robot without adjusting the video model at all. When the backbone is fine-tuned alongside the decoder, the lab claims state-of-the-art success rates.

Two further numbers matter if they hold. The first is data. Black Forest Labs says FLUX-mimic can be fine-tuned for a specific manipulation task "with as little as 30 minutes of robot data, where prior approaches have required 30 or more hours" — a roughly sixty-fold reduction in the most expensive input in robot learning. mimic CTO Elvis Nava framed the problem this way: "The hardest part of robotics is data. Every new task normally means hours of a robot repeating itself. Because FLUX-mimic is built on top of frontier video models that already understand how the physical world behaves, it picks up a new task in minutes, not days."

The second is latency, which is where most impressive robotics demos quietly fail. The company says the FLUX-mimic backbone can be optimised to run from input to world representation in under 80ms on a single NVIDIA RTX 5090, and that mimic's full deployment stack — action decoder, sensor-to-actuator plumbing, and overlapping prediction with execution — yields "a self-contained robot system with reaction times of 101ms." That is roughly human visual reaction time, on one consumer-class GPU.

There is also a behavioural claim worth flagging because it is the kind of thing a world model should buy you: the lab says FLUX-mimic recovers from failures it was never shown. "A robot that misses a grasp corrects itself, grasps again, and completes the task." No demonstration set covers every way a task can go wrong; recovery that was never demonstrated has to come from a model that already knows how the world behaves.

What Audi says

The deployment partner is the part that separates this from a research post. Audi's Production Lab has been running FLUX-mimic on exactly the work that has resisted automation for decades — kitting parts into structured trays, inserting electronic control units into tight-fitting fixtures, and handling soft, flexible materials like seals and cables. The reason those tasks stayed manual is economic rather than technical: the variant diversity of premium car production makes conventionally programmed robot cells too costly to re-engineer for each case.

Christoph Schneider of Audi Production Lab is quoted directly: "In partnership with mimic, Audi has been testing and deploying FLUX-mimic. We have seen these robots solve complex soft body manipulation work that would have been simply impossible with conventional robotics."

Note the precise wording — testing and deploying, via the Production Lab, which is Audi's pre-production piloting unit rather than its main assembly line. This is a real industrial pilot with a named Audi contact attached, not a lab demo. It is not yet evidence of production-scale rollout, and nobody involved claims it is.

The numbers you should not take on trust

Every performance figure above comes from Black Forest Labs. The company is unusually upfront about this: it describes its video evaluations as "preliminary," expects them to improve, and says it "will publish full FLUX 3 benchmark results and methodology alongside broader availability" — which is to say, the methodology is not published yet.

The video comparisons are human-preference win rates from the vendor's own harness. FLUX 3 was reportedly preferred over Luma Ray 3.2 in 93% of comparisons, Runway Gen-4.5 in 77%, Grok Imagine Video in up to 69%, Kling v3 Pro in 60%, Happy Horse v1 in 59%, Happy Horse 1.1 in 57%, and Seedance 2.0 and Gemini Omni Flash in 52% each. Vendor-run preference tests against competitors are the least independently checkable class of benchmark there is, and a 52% result is a coin flip. Treat the whole set as a claim awaiting replication.

The robotics benchmarks carry the same caveat, with an additional one: video-action models are an existing research direction, not something invented here. Black Forest Labs itself cites prior work reporting up to 10x sample efficiency for video-action models over vision-language-action models. What is new is a frontier generative-video lab shipping one into a car factory, not the idea.

The open-weights thread

FLUX 3 Video and FLUX 3 Action are in early access now. FLUX 3 Image rolls out in the coming weeks. And FLUX 3 Dev — an open-weight multimodal backbone covering both content generation and action prediction — is promised later this year.

That promise lands in a specific political week. Black Forest Labs' stated rationale is operational rather than ideological: open weights make "secure, low-latency local deployment possible for applications like robotic control systems," and let teams adapt the model to their own data. If your robot's controller has to answer in 101ms on a factory floor, a remote API is not a serious option.

It is a pointed argument to make right now, and Black Forest Labs is making it in two places at once. Washington spent this week arguing about whether to restrict open-weight models at all: nearly 200 startups wrote to the administration on Wednesday opposing a ban on Chinese open-weight systems, and on Friday a coalition of 25 companies and organisations — Nvidia, Microsoft, Meta, IBM, Dell, Palantir, Hugging Face, Mistral, Mozilla, the Linux Foundation and Y Combinator among them — published a joint letter titled "Open Weights and American AI Leadership". Black Forest Labs signed it. Its co-founder and CEO has separately urged G7 leaders to back open innovation.

A European lab announcing that it will open-weight a model which physically operates machinery adds a category to that debate that the current framing — mostly about chatbots and national security — does not really have a slot for. The letter's own risk section concedes that once weights are released they are "beyond the original developer's control, and modified versions are difficult to trace or reverse." That concession reads differently when the model in question can drive a robot arm.

For context on where the open-weight field stood before this, see our mid-2026 open-weight scoreboard.

What to watch

Three things will tell you whether FLUX 3 is what the company says it is.

Published methodology. Black Forest Labs has committed to releasing full benchmark results and methodology at broader availability. Until then the 93% and the 30-minutes figures are marketing, however plausible.

Independent robot replication. The frozen-backbone result is the load-bearing claim. If other groups can bolt an action decoder onto a general video model and beat purpose-built vision-language-action systems, that reorganises a lot of robotics research. If it only works with mimic's deployment stack and mimic's hardware, it is a good product rather than a general result.

Whether FLUX 3 Dev actually ships with action prediction. Open-weighting a video generator is now routine. Open-weighting a backbone that has been demonstrated driving industrial robots is a materially different proposition, and it is the one that will test both the company's stated commitment and the regulatory mood it is walking into.

Black Forest Labs is valued at $3.25 billion and has raised more than $450 million from investors including a16z, NVIDIA, Salesforce Ventures, Adobe Ventures, General Catalyst, Canva and Deutsche Telekom's T.Capital. The 100-person team works out of Freiburg and San Francisco. It launched from stealth in 2024.