XNews

FLUX 3 Multimodal Model Release

FLUX 3 Multimodal Model Release

Black Forest Labs released FLUX 3, a multimodal model that unifies image, video, and audio within a single flow architecture, aiming to learn how the world works. The model is positioned as a 'real-world model' rather than just a generative tool, with three product lines: Video (early access, up to 20 seconds with native audio, supports text/image/video reference, keyframes, multilingual dialogue, chaining), Image (coming weeks, synthesis and editing with complex prompts and multilingual text), and Action (partner channel, native action prediction, already used in FLUX-mimic for robotics). The model will be open-sourced as Flux 3 Dev, with API access opening in the coming weeks. Features include video reference generation, video/audio extension, keyframe control, smart multi-shot editing, text animation, and image output.

Key insights

FLUX 3 represents a paradigm shift from single-modality generation to a unified world model that learns cross-modal constraints

Black Forest Labs designed FLUX 3 to learn from the mutual constraints across image, video, audio, and language—such as sound matching impact and motion obeying mass—arguing that this joint training yields a more coherent world representation than training on each modality separately. The same base model connects to content generation and embodied action prediction, positioning it as a 'real-world model' rather than a narrow generative tool.

The open-source release and broad feature set aim to make advanced multimodal generation accessible to developers and creators

Flux 3 Dev will be open-sourced, with early access available now and API access in the coming weeks. The model supports a wide range of features including video reference generation, keyframe control, multilingual dialogue, smart multi-shot editing, and image output, catering to both creative and technical use cases.

The inclusion of action prediction and robotics applications signals an expansion beyond media generation into embodied AI

FLUX 3's Action product line enables native action prediction and can use the video backbone as a dynamics foundation for fine-tuning in robotics. The FLUX-mimic project, applied in mimic robotics and an Audi scenario, demonstrates early real-world use of this capability.

Source posts