Black Forest Labs launches FLUX 3 for 20-second videos

Black Forest Labs released FLUX 3 in early access on Thursday. The model generates video clips up to 20 seconds long with audio synchronized to on-screen events. The same model backbone is used in FLUX-mimic, a robotics system currently under trial by Audi on a production line.

FLUX 3 was trained on images, video and audio together in a single multimodal model. The company is offering Video and Action capabilities first through APIs and select partners, with Image generation to be made available in the coming weeks. An open-weight “Dev” version intended for local use is scheduled for later in 2026.

In internal preference tests, human reviewers chose FLUX 3 over Runway Gen-4.5 in 77% of pairings and over Luma Ray 3.2 in 93% of pairings. The model was preferred to Gemini Omni and Seedance in 52% of comparisons. Black Forest Labs described the evaluation as a head-to-head preference test in which reviewers picked the clip they found more convincing.

Robin Rombach, co-founder and chief executive, argued that a model trained on video can learn physical properties such as weight, contact and timing in addition to visual appearance. That training approach is what Black Forest Labs applied when creating FLUX-mimic with Zurich-based mimic robotics.

FLUX-mimic adds a lightweight decoder to the FLUX 3 backbone that converts the model’s video-prediction outputs into robot motion commands. Audi is using the system for tasks such as fitting flexible door seals, a type of soft-body manipulation that the company reported conventional automation has struggled to perform. Mimic co-founder Stephan-Daniel Gravert described Audi as the manufacturing partner they built the system for. Christoph Schneider at Audi said the robots now “solve complex soft-body manipulation work” that older machines could not handle.

Black Forest Labs reported that the full FLUX-mimic system responds in about 101 milliseconds. FLUX 3 also produces still images and supports a range of visual styles beyond photorealism.

The company was founded in August 2024 by researchers who had worked on the original Stable Diffusion models. Earlier FLUX releases included open-source Flux Dev and Schnell models and a non-open FLUX 1.1 Pro. FLUX.2, released in November 2025, did not match the earlier open-source popularity. In late 2025, Alibaba released Z-Image Turbo, which matched image quality on lower-end hardware.

Video and Action features of FLUX 3 are available now to partners and via API access. Image generation will be rolled out in the coming weeks, and the open-weight Dev release is expected later in 2026.