FLUX 3 Video exits early access onto Runway and Krea
Black Forest Labs' 20-second video model with native audio opened to the public on Runway, Krea and Replicate after a gated rollout.
From Discord waitlist to public endpoints
Black Forest Labs' FLUX 3 Video became broadly available on August 4, appearing the same day on Krea, Replicate and Runway. The model had been announced on July 23 but shipped only through a gated early-access program, with production use restricted and no public API or third-party endpoints; as of late July neither Replicate nor fal.ai carried it. Tuesday's simultaneous listings mark the point at which ordinary creators can actually generate with it.
What the model does
FLUX 3 is a single multimodal flow model covering image, video, audio and robot action prediction from one backbone, built on BFL's Self-Flow approach to unifying generation and understanding. The video component produces clips of up to 20 seconds at 720p with audio synthesised in the same pass rather than dubbed afterwards, and supports text-to-video, image-to-video, video-to-video from a reference clip, keyframe-controlled transitions, and generative continuation of an existing clip with its audio. BFL has emphasised facial expression fidelity, sound tied to on-screen physical events, and multilingual dialogue. 1080p output and a FLUX 3 Image release are still pending, as is the open-weight FLUX 3 Dev variant the company has said will follow later this year.
Runway's listing is the notable one: the platform is now serving a rival lab's video-plus-audio model alongside its own generation stack, a continuation of the aggregator posture it has adopted in recent weeks.
Why it matters
Generative video's competitive axis has moved from clip length to whether picture and sound are produced together, and FLUX 3 arrives as one of the few models doing that natively at 20 seconds. Broad availability also settles a practical question for studios and agencies that could not commit to an early-access model with restricted commercial terms. The action-prediction head is the longer-term story: BFL is training the same weights that generate footage to also predict physical actions, positioning a company known for image generation inside the world-model and robotics race.