Black Forest Labs launches FLUX 3 capable of generating images and 20-second video with audio — but in…

Black Forest Labs, the Freiburg, Germany-based AI lab also known as BFL, has launched FLUX 3, expanding its FLUX family beyond image generation. The model is positioned as a multimodal frontier system that can understand and generate images, or produce combined audio and video clips up to 20 seconds long from a single prompt. The company is starting with a limited release rather than a broad public rollout.

Technically, BFL describes FLUX 3 as jointly trained across modalities instead of bolting together separate image, video, and audio stacks. The same underlying architecture is also meant to extend into robotic vision and actions, so perception and generation share one training and inference design rather than a pipeline of specialist models. That joint-training choice is the core product mechanic: one prompt in, image or short audiovisual clip out, with robotics framed as a further use of the same backbone.

For engineers and builders, a single model that covers stills plus short video with audio cuts the usual glue work of chaining an image model, a video model, and a separate audio or speech stack. Limited release means early access will matter more than marketing claims: who gets API or product access, what latency and resolution look like in practice, and whether the 20-second audiovisual path is stable enough for product prototypes will decide whether FLUX 3 is usable in real pipelines or only as a research preview.

In market terms, FLUX has been known mainly as an image-generation line; FLUX 3 is BFL’s move into the same multimodal, short-form video-and-audio space that frontier labs are racing to own. Joint training plus a robotics path signals that BFL is competing on unified world-model style systems, not only on still-image quality. Limited availability also fits a common pattern: ship the flagship claim first, control capacity and safety surface while demand and quality are still being proven.

What to watch next is how the limited release expands—who is allowed in, what the generation surface exposes (images only vs full 20-second A/V), and whether the robotic vision and action story shows up as concrete APIs or demos rather than roadmap language. Until that access widens, the practical takeaway is to treat FLUX 3 as a capability announcement with a hard cap on clip length and a deliberately gated rollout, not as a drop-in replacement for existing image-only FLUX workflows.