Black Forest Labs (BFL) has released FLUX 3, a multimodal foundation model designed to learn from images, video and audio within a single architecture. The company describes it as the first FLUX model to deliver video, audio and action prediction from one shared set of weights.

The BFL research team says the model is based on the view that no single modality provides a complete description of the world. Images capture spatial structure at a single moment, video adds time and physical dynamics, and audio can reveal causal relationships between mechanical events and sound. In BFL’s framing, each modality is a lossy projection of the same underlying reality.

By training across these modalities at the same time, BFL says the data sources constrain one another: sound must correspond to impact, and motion must be consistent with mass. The research team calls FLUX 3 its first model built entirely around that principle, with the shared-weight design also making the release relevant beyond media generation because BFL is applying the same backbone to robot action prediction.

Self-Flow as the underlying method

FLUX 3 builds on Self-Flow, BFL’s method for aligning multimodal generation and understanding in one architecture. Self-Flow combines a flow matching objective with a self-supervised feature reconstruction objective.

The reference implementation on GitHub is released under Apache-2.0 and uses SiT-XL/2 with per-token timestep conditioning. It trains with a 25% per-token mask ratio and uses self-distillation from an EMA teacher at layer 20 to a student at layer 8.

That released checkpoint is an ImageNet 256×256 research model, not FLUX 3. BFL states that it “significantly scaled up compute and data resources” using the same approach to train FLUX 3 simultaneously across video, images and audio. Self-Flow was introduced in March 2026, so the method itself is not new to this release; BFL identifies the new element as scale.

FLUX 3 Video capabilities

FLUX 3 Video can generate clips up to 20 seconds long in a single generation and includes native audio. Supported modes include text-to-video, image-to-video, video-to-video from a reference clip, keyframe-to-video for controlled transitions, and generative video-audio continuation from input video and audio.

BFL also lists multilingual dialogue, agentic chaining of clips into multi-shot sequences, and strong typography generation with animated designs among the system’s capabilities. The team reports particular strength in generating human facial expressions and in associating sounds with physical events, two areas where temporal consistency and audio-video alignment are central to user perception of generated clips.

According to BFL, video prediction accounts for more than 95% of the training compute, while audio represents less than 0.5% of tokens. The company also says the same backbone powers FLUX-mimic, a robot policy that runs in under 80 ms on one RTX 5090.

Preliminary performance results

BFL published preliminary human preference results using 10-second text-to-video clips at 720p with audio. In those comparisons, FLUX 3 was preferred over Luma Ray 3.2 in 93% of cases and over Runway Gen-4.5 in 77%.

Against Grok Imagine Video, BFL reported a preference rate of up to 69%. The reported figures were 60% against Kling v3 Pro, 59% against Happy Horse v1 and 57% against Happy Horse 1.1. Against Seedance 2.0 and Gemini Omni Flash, the reported result was 52%, which the source described as close to a coin flip.

Access remains gated. BFL says Video and Action are in early access, Image access will follow, and open weights will come last. That staged rollout means outside evaluation will depend on when broader access, model details and weights become available.

BFL published additional materials through the FLUX 3 announcement, the FLUX 3 x mimic technical post and the Self-Flow paper.