Model Releases
Black Forest Labs Releases FLUX 3 Multimodal Foundation Model
Black Forest Labs shifts from diffusion to unified world modeling with the release of FLUX 3, a multimodal architecture integrating image, video, audio, and action prediction.
July 23, 2026
Model ReleaseFLUX 3
Introduction
On July 23, 2026, Black Forest Labs (BFL) marked a historic pivot in the evolution of generative AI. Moving beyond the specialized diffusion models that defined the FLUX.1 and FLUX.2 eras, the company has unveiled FLUX 3. This is not merely an incremental update or a model with video capabilities 'bolted on' to an image generator; it is a fundamental shift toward a unified multimodal foundation model.
FLUX 3 is positioned as a 'model of the world' rather than a collection of disparate modalities. By jointly learning from images, videos, and audio within a single, cohesive architecture, BFL aims to capture the underlying physics and temporal logic of the real world. This release signals the beginning of the 'Real World Models' era, where visual intelligence is grounded in a holistic understanding of sensory data.
- Announced: July 23, 2026
- Architecture: Unified Multimodal Flow Model
- Core Modalities: Image, Video, Audio, and Action Prediction
- Status: Early Access API available
Key Features & Architecture
The technical backbone of FLUX 3 represents a departure from traditional latent diffusion. BFL describes the architecture as 'multimodal flow models,' designed to serve as the backbone of visual intelligence. Instead of processing modalities in silos, the model utilizes a unified latent space where audio cues, visual frames, and physical action tokens are processed through a shared transformer-based flow mechanism.
This unified approach allows for unprecedented cross-modal reasoning. For instance, the model can predict how a specific sound (audio) should correspond to a visual movement (video) or a physical manipulation (action). This architecture is specifically optimized for high-dimensional temporal consistency, making it a prime candidate for both high-fidelity media generation and robotics integration.
- Unified Architecture: Single backbone for all modalities
- Multimodal Flow: Advanced flow-matching for temporal and spatial coherence
- Action Integration: Native support for physical action prediction tokens
- Scalable Family: Includes specialized variants like Image Klein and Action models
Performance & Benchmarks
FLUX 3 demonstrates a significant leap in cross-modal alignment compared to the FLUX.2 lineage. In internal benchmarks focusing on temporal consistency and audio-visual synchronization, FLUX 3 outperformed previous iterations by a margin of 40% in video-to-audio coherence scores. The model's ability to maintain structural integrity in long-form video generation sets a new industry standard for foundation models.