Black Forest Labs announces the multimodal FLUX 3

Background

Black Forest Labs (BFL) had first published Self-Flow, a research technique for aligning understanding and generation across modalities, in March 2026. FLUX 3 was built on that research.^[1]

What happened

On 23 July 2026 Black Forest Labs announced FLUX 3, a model family for images, video, audio and actions.^[1]

FLUX 3 Video and FLUX 3 Action launched in gated early access. FLUX 3 Image and an open-weight FLUX 3 Dev were promised for later.^[1]

Key details

FLUX 3 Video generates clips of up to 20 seconds with native audio.^[1][2] On 4 August 2026, BFL made FLUX 3 Video generally available through the BFL API.^[2]

CEO Robin Rombach said: “You can’t cheat reality.”^[1]

Impact

At the time of the announcement, FLUX 3 Image and FLUX 3 Dev were still planned, and FLUX 3 Video and FLUX 3 Action were in gated early access. FLUX 3 Video became generally available on 4 August 2026.^[1][2]

Several terms in this announcement need explaining. A multimodal model handles more than one kind of data, such as images, video and audio, within one model family. Gated early access means that only approved users can try a model, and no weights are downloaded. An open-weight release, by contrast, publishes the model's parameters so that anyone with suitable hardware can run it, subject to its license.

The announcement therefore described a staged rollout. Some parts were already in limited use at the time, while others were still planned, and each part's availability depended on its own release schedule.

For readers outside the field, the key idea behind a multimodal family is shared design. Instead of separate systems for pictures, video and sound, one design can be trained to read and produce several kinds of media, which makes it easier to relate them to one another.