Black Forest Labs Stopped Being an Image Company. FLUX 3 Predicts Video, Audio, and Robot Motion From One Backbone.

TL;DR

On July 23, Black Forest Labs, the German lab founded by the researchers behind latent diffusion and Stable Diffusion, announced FLUX 3. It is not another image model. It is a single multimodal foundation model that jointly learns images, video, and audio in one architecture, generates video with synchronized audio up to 20 seconds in a single pass, and predicts robot actions from the same backbone. FLUX 3 Video and FLUX 3 Action are in early access now; FLUX 3 Image lands "in the following weeks." The part you can download is still a promise: an open-weight FLUX 3 Dev is planned for later in 2026. The lab that taught a generation of homelabbers to run image models on a 4090 just repositioned itself as a world-model company, and the demo it picked was a robot arm at Audi.

What actually shipped, and what did not

Read the availability line before you plan a sprint around this. FLUX 3 Video, with optional native audio, is in early access. FLUX 3 Action is in early access. FLUX 3 Image, the thing most people associate with the FLUX name, is not out yet and is described as rolling out in the coming weeks, with general availability after that. Named early testers include Canva, Krea, Picsart, Magnific, and Burda.

So the announcement is a capability disclosure, not a launch you can build on today. That distinction has been getting load-bearing lately, and it is worth holding onto.

Self-Flow: one model, four jobs

The architectural claim is the interesting bit. Black Forest Labs calls its method Self-Flow, described as an approach for "efficiently aligning multimodal generation and understanding within the same underlying architecture." In plain terms: instead of bolting a video model, an audio model, and a policy network together with glue code, one model learns to both produce and interpret every modality, and FLUX 3 scales that recipe up with more compute and data.

Why would a video model be any good at moving a robot arm? Because predicting the next frame and predicting the next joint angle are close to the same problem in different clothes. A model that can guess how a hand closing around a rubber seal will look one second from now has already absorbed most of the physics; what remains is translating "the pixels move like this" into "the actuators move like that." It is the difference between someone who has watched ten thousand hours of cooking shows and someone who has never seen a kitchen. Both still have to pick up the knife, but only one of them is starting from zero.

CEO Robin Rombach put the thesis less gently: "You can't cheat reality. A model that only learns images can only generate images."

The video numbers, with the usual asterisk

FLUX 3 Video does text-to-video, image-to-video, video-to-video, and keyframe-to-video, with audio generated alongside the picture, including multilingual dialogue. Clips run up to 20 seconds with audio in a single generation. The published preference evaluations were run at 720p on 10-second clips, which is a narrower slice than "20 seconds" implies.

Those evaluations are human preference win rates, measured and published by Black Forest Labs itself:

Read that chart the way the lab wrote it. Against Luma Ray 3.2 and Runway Gen-4.5, FLUX 3 wins comfortably. Against Grok Imagine Video and Kling v3 Pro it wins clearly. Against ByteDance's Seedance 2.0 and Google's Gemini Omni Flash it lands at 52%, which is a statistical shrug. The lab says so itself, noting that "evaluations are early." Two models named Happy Horse also appear on the comparison list at 57% and 59%, which tells you roughly everything about the current state of video-model naming.

The robot half is the actual news

FLUX 3 Action is where this stops being a creative-tools story. Black Forest Labs built FLUX-mimic with mimic robotics, an ETH Zurich spinout working on dexterous manipulation, by combining the FLUX 3 video backbone with mimic's robot-learning stack. Audi is testing it on real production and logistics tasks involving soft-body manipulation, the flexible-material problems that conventional industrial automation has historically been terrible at.

The claim that matters is the data efficiency. Per the companies, a new manipulation task can be fine-tuned with as little as 30 minutes of robot data, against 30 or more hours previously.

Elvis Nava, mimic's CTO, framed it in the currency that matters on a factory floor: "Every new task normally means hours of a robot repeating itself... it picks up a new task in minutes, not days."

Treat that as a vendor claim until somebody outside the two companies reproduces it. But the direction is the point. If a generative video model transfers into manipulation policies this cheaply, the scarce resource in robotics shifts from teleoperated demonstrations to pretraining video, and video is the one thing the internet has an infinite supply of.

Where the open weights went

Here is the part this audience will care about most, and it is unresolved. FLUX has been the default local image stack for two years precisely because the weights showed up. FLUX.2 [dev] is a 32B flow-matching transformer sitting on Hugging Face under a non-commercial license, with an official inference repo and an fp8 path for consumer RTX cards.

For FLUX 3, the commitment is real but deferred: open-weight access to a multimodal backbone covering video, audio, image, and action prediction, shipping as FLUX 3 Dev later in 2026. No parameter count, no license terms, no date. If it lands as described, an open-weight model that generates video with audio and also predicts robot actions would be a genuinely new object in the local-AI ecosystem. If it slips, FLUX 3 is an API you rent.

The caveats

• Every performance number here is vendor-published. No independent arena results exist yet.
• The preference evals were run at 720p on 10-second clips, not at the headline 20-second length.
• FLUX 3 Image, the flagship modality for most existing FLUX users, is not available.
• The Audi work is described as testing, not deployment.
• The 30-minutes-of-robot-data figure comes from the two companies with a product to sell.
• Full technical details are promised later, so Self-Flow is currently a name and a claim, not a paper you can check.

Key Takeaways

• Black Forest Labs announced FLUX 3 on July 23, 2026: one multimodal foundation model trained jointly on images, video, audio, and action, built on a method it calls Self-Flow.
• FLUX 3 Video generates up to 20 seconds with native synchronized audio in a single pass, supporting text, image, video, and keyframe conditioning.
• Vendor preference tests put it well ahead of Luma Ray 3.2 (93%) and Runway Gen-4.5 (77%), but at 52% parity against Seedance 2.0 and Gemini Omni Flash.
• FLUX-mimic, built with ETH spinout mimic robotics, adapts the video backbone to robot manipulation and is being tested at Audi on soft-body tasks, with a claimed 30 minutes of robot data per new task versus 30-plus hours before.
• Video and Action are in early access now; Image ships in the coming weeks; the open-weight FLUX 3 Dev is promised for later in 2026 with no specs or date.
• The lab is valued at $3.25 billion after a $300 million Series B and is now explicitly betting that visual generation and physical action are the same problem.

Sources: Black Forest Labs: FLUX 3, GlobeNewswire press release, Decrypt, Black Forest Labs: our $300M Series B, Black Forest Labs: FLUX.2, FLUX.2 [dev] model card