# Black Forest Labs launches FLUX 3 across video, audio and robot actions > Source: > Published: 2026-07-23 16:10:06+00:00 Robin Rombach's [Black Forest Labs (@bfl_ai)](https://x.com/bfl_ai/status/2080308988961554582) introduced FLUX 3 on July 23rd, 2026, extending its image-generation technology into video with native audio and using the same underlying model architecture to predict actions for robots. The immediate release is narrower than the four-part launch suggests. [FLUX 3 Video](https://bfl.ai/models/flux-3?utm_source=x&utm_medium=social&utm_campaign=flux3_launch) is available through a selective early access program, while image generation, action prediction and open weights are scheduled to arrive over the coming weeks and months. Black Forest Labs has not published pricing, a parameter count, a general availability date or a complete technical report. Rombach built Black Forest Labs with researchers whose work shaped the previous generation of generative media. Rombach, Andreas Blattmann and Patrick Esser were authors of the influential [latent diffusion paper](https://arxiv.org/abs/2112.10752), the research foundation that later supported Stable Diffusion. FLUX 3 carries that work beyond still images by training jointly on images, video and audio, rather than connecting separately trained systems at inference time. In its [technical launch post](https://bfl.ai/blog/flux-3), Black Forest Labs describes FLUX 3 as its first model built entirely around a shared representation of the physical world. The stated thesis is that images provide spatial structure, video adds motion and physical behavior, and audio adds another record of cause and effect. Language connects those signals to instructions. That architecture allows FLUX 3 Video to produce clips with synchronized audio from text, images or existing video. Black Forest Labs says the model can generate clips up to 20 seconds in one pass and supports text-to-video, image animation, video-to-video transformation, clip continuation, keyframe transitions and multilingual dialogue. The model can also chain clips into longer sequences while using visual references to preserve characters. Black Forest Labs evaluated 10-second, 720p clips with audio and reported preference rates ranging from 52% against Seedance 2.0 and Gemini Omni Flash to 77% against Runway Gen-4.5 and 93% against Luma Ray 3.2. Those are preliminary, company-run results. Black Forest Labs did not publish the number of comparisons, the evaluator pool or a reproducible benchmark package with the launch, limiting what the percentages establish. The more consequential part of Rombach's bet sits outside media generation. Black Forest Labs says video training consumed more than 95% of FLUX 3's compute because predicting video requires the model to learn motion, contact, weight and cause and effect. The lab is trying to reuse that expensive representation for robot control, where actions can be treated as another sequence tied to visual observations. Black Forest Labs gave [mimic robotics](https://www.mimicrobotics.com/) early access to the model and jointly developed FLUX-mimic, a video-action system for dexterous industrial manipulation. In a [separate technical post](https://bfl.ai/blog/flux-3-mimic), Black Forest Labs says FLUX-mimic places a lightweight action decoder over features extracted from FLUX 3's video-prediction path. Black Forest Labs says the system has been used for factory tasks including placing parts into trays, inserting electronic control units into fixtures, assembling components and manipulating flexible seals and cables. Audi Production Lab said it has tested and deployed FLUX-mimic for complex soft-body manipulation work. Black Forest Labs reported a backbone latency below 80 milliseconds on a single Nvidia RTX 5090 and an end-to-end robot reaction time of 101 milliseconds, though the launch material does not provide an independent performance audit. The robotics work gives FLUX 3 a route into industrial automation while its video product enters a crowded generation market. Black Forest Labs is positioning the same training run as infrastructure for creative software and physical machines, spreading the cost of a large video model across two distinct businesses. Black Forest Labs has capital to pursue that expansion. In December 2025, Black Forest Labs [raised a $300 million Series B](https://bfl.ai/blog/our-300m-series-b) at a $3.25 billion post-money valuation, co-led by Salesforce Ventures and Anjney Midha's AMP. The round also included Temasek, Bain Capital Ventures, Air Street Capital, Visionaries Club, Canva and Figma Ventures, alongside existing backers including Andreessen Horowitz, Nvidia and General Catalyst. For now, access remains controlled. The [early access form](https://tally.so/r/44d9NX) says participation is free and prioritized according to use case and fit. Black Forest Labs plans to release video and audio through APIs and private weights, action prediction through selected partners, image generation through APIs and private weights, and an open-weight multimodal backbone called FLUX 3 Dev.