FLUX 3 Released: Black Forest Labs Unveils 'Real World Model'

Black Forest Labs released Flux 3 on July 23, 2026, a multimodal foundation model that simultaneously learns images, video, and audio. Flux 3 is the first to support generating videos with native audio, up to 20 seconds. BFL believes that a single modality cannot fully capture reality, and joint training of the three can fill information gaps.

Flux 3 supports text-to-video, image-to-video, video-to-video, keyframe transitions, multilingual dialogue, and smart editing (stitching clips into multi-shot sequences). BFL states the model excels at matching human facial expressions and sounds with physical events.

The model is based on the Self-Flow method, using a multimodal Transformer that converts images, video, audio, and motion into a shared internal representation via dedicated encoders and decoders, then converts back to output. The motion component provides a foundation for robotics applications.

BFL conducted early evaluations using 10-second 720p videos for user preference tests. Flux 3 achieved a 93% preference rate over Luma Ray 3.2, 77% over Runway Gen-4.5, and 69% over Grok Imagine Video.

The advantage narrows against stronger competitors: 60% over Kling v3 Pro, 59% over Happy Horse v1, 57% over Happy Horse 1.1, and 52% over both Seedance 2.0 and Gemini Omni Flash. BFL notes the results are preliminary and no independent tests have been conducted.

BFL plans a phased release: Flux 3 Video is already available, Flux 3 Image will enter early access within weeks, and motion prediction features are initially limited to partners.

BFL will release an open-source weights version, Flux 3 Dev, but currently it is only accessible via early access application. The API will open in the coming weeks, followed by the open-source version.

BFL collaborated with Mimic Robotics to develop Flux-mimic, a video motion model that has been tested in Audi production tasks. This demonstrates the potential of Flux 3's unified learning framework in robotics.