Introduction

The gates are open, sunshine. On August 4, 2026, Black Forest Labs moved FLUX 3 Video from early-access whispers into general availability on the BFL API and select partners — clips up to 20 seconds, HD natively, Full HD via upscaling, and native audio baked into the same generation. Dialogue, Foley, ambience: not a second model duct-taped on after the fact.

If you’ve been waiting for a video model that feels like real-world visual intelligence instead of a silent slideshow with a soundtrack bolted on later… this is your cue to press record.

What FLUX 3 Video Is

FLUX 3, announced July 23, 2026, is BFL’s multimodal foundation model — jointly trained on images, video, and audio under their Self-Flow approach, with a tagline that leans into real-world visual intelligence. Video is one face of that same model family, not a cousin living in another repo.

FLUX 3 Video generates picture and sound together: speech with lip-sync, sound effects, room tone — the whole sonic weather system of a shot. Modes you’ll meet in the wild include:

  • Text-to-video — from a simple beat or a director’s paragraph
  • Image-to-video / keyframes — start frame, end frame, or a handful of anchors
  • Video-to-video reference and generative video-audio continuation
  • Keyframe-to-video and multi-shot / agentic chaining inside one clip
  • Multilingual dialogue + lip-sync, broad styles and aspect ratios, plus surprisingly strong typography in-scene

Think of it as a tiny film unit that already knows the boom mic is part of the camera.

July → August: The GA Timeline

July 23: BFL unveils FLUX 3 and teases the multimodal stack — video with native audio, early preference numbers, and a roadmap that still includes richer reference mixes, FLUX 3 Image, and open-weight Dev later in 2026.

August 4: FLUX 3 Video, Part 1: Generation — the initial generation stack goes generally available via the BFL API and select partners. Not “email us for a waitlist forever.” Available.

Before release, BFL says they worked with third-party partner Cinder to evaluate mitigations across modalities, including NCII and CSAM risk — the unglamorous safety work that still belongs in the sunshine of a launch note.

What You Can Make Now

GA specs creators can actually ship with:

  • Clips up to ~20 seconds
  • HD 720p native, Full HD via upscaling
  • Draft Mode for cheap, fast previews — then promote the winner to full quality with matching subjects, composition, and motion
  • Video continuation from up to about 4 seconds of existing video plus audio (movement, camera, dialogue, and sound across the seam)
  • Multi-shot sequences inside a single generation
  • Many languages and dialects (English varieties, Chinese, Spanish, French, German, Japanese, and more) with lip-sync that aims to feel lived-in

On the API side, generation rides the flux-3-video endpoint with modes like t2v, i2v, and v2v (continuation). Pricing is usage-based and can change — as of the public docs, examples look like roughly $0.17/s HD and $0.29/s FHD for full text/image-to-video renders, with draft cheaper (on the order of ~$0.06/s), and continuation billed higher. Always confirm live numbers on the official docs rather than tattooing a blog post to your invoice:

https://docs.bfl.ai/flux_3https://bfl.ai/blog/flux-3-videohttps://bfl.ai/models/flux-3

Want a playground wander before you wire production? Peek at BFL’s site — and if you’re browsing ArtRealmAI’s tool shelf, gen.html is a friendly place to keep comparing video models side by side.

Pro Tip

Treat Draft Mode like a sunny table read: spend your cheap takes on prompt, camera language, and audio beats, then spend the full-quality render only on the take you’d actually show a client. Continuation loves a clean ~4-second handoff — leave the seam a little breathing room so motion and dialogue can cross without a jump cut scare.

How It Stacks Up

BFL’s July announcement shared vendor, preliminary human preference shares for FLUX 3 Video versus several peers — useful as a snapshot of their early evals, not as eternal gospel:

  • Preferred over Grok Imagine Video ~69%
  • vs Kling v3 Pro ~60%
  • vs Seedance 2.0 / Gemini Omni Flash ~52%
  • vs Runway Gen-4.5 ~77%
  • vs Luma Ray 3.2 ~93%

The August 4 GA post says capabilities progressed further since that announcement, with internal raters still preferring FLUX 3 for text-to-video and image-to-video overall (and noting a tie with Seedance 2.0 on i2v in their framing). Keep independent benches on your radar — sunshine ages best when more than one weather station agrees.

What’s Next

The roadmap is still deliciously unfinished:

  • Richer reference mixes (image + video + audio together)
  • FLUX 3 Image early access “following weeks” from the July post
  • FLUX 3 Dev open-weight later in 2026
  • Action / FLUX-mimic threads with mimic robotics (including Audi production-style tests mentioned around the family announcement)

Generation GA is Part 1 — controllability and new modality combos are queued behind it.

Conclusion

FLUX 3 Video’s GA moment is simple and bright: 20-second clips, native audio in the same pass, Draft-to-final iteration, continuation that remembers sound, and an API you can actually call without a secret handshake. The forest just got a lot more cinematic — dialogue and all.

Go make something that sounds as good as it looks.

—Aurelia ♡