FLUX 3

What it is

FLUX 3 is a multimodal model family from Black Forest Labs (BFL). Multimodal means that one family handles several kinds of media rather than only still pictures. BFL announced it on 23 July 2026, and it covers images, video, audio and action prediction. It is the first BFL family to go beyond still images.^[1]

Origins

FLUX 3 builds on Self-Flow, a research technique that BFL first published in March 2026.^[1] Self-Flow is used to align understanding and generation across different modalities inside a single architecture, which means one model design covers reading and producing several kinds of media.^[1]

Variants

FLUX 3 comes in several variants: FLUX 3 Video, FLUX 3 Image, and FLUX 3 Action, which handles robotic vision and action prediction.^[1] BFL also planned an open-weight model, FLUX 3 Dev.^[1] At launch, FLUX 3 Video and FLUX 3 Action were in gated early access, meaning access was limited and there were no downloadable weights.^[1]

FLUX 3 Video

FLUX 3 Video generates clips of up to 20 seconds with native audio. The audio can include multilingual dialogue with lip-sync, sound effects and ambient sound.^[2] Video renders natively at 720p, and 1080p output is produced by upscaling.^[2]

The video model has several modes: text-to-video, image-to-video, keyframes, and continuation from up to four seconds of input video.^[2] Keyframes are chosen frames that set the start or important points of a shot, and the rest of the clip is filled in around them, which is a general description of the technique.

Benchmarks

In BFL’s own human-preference evaluation, FLUX 3 Video led text-to-video with an Elo score of 1135. On image-to-video it tied Seedance 2.0.^[2] These results come from BFL’s evaluation, so they reflect the company’s own testing.

Availability

On 4 August 2026 BFL made FLUX 3 Video generally available through the BFL API and select partners.^[2]

Robotics and open weights

A FLUX 3 x mimic model for robotics was announced together with Mimic Robotics.^[3] FLUX 3 Klein was released with open weights, about 5.6 billion parameters, on 15 October 2026.^[3]