FLUX 3 Video: What Black Forest Labs' 20-Second Model Means for Creators

If you have spent an afternoon stitching three 5-second AI clips into one ad, you already know the tax. Faces drift. Lighting jumps. Audio from three tools never quite lines up. On July 23, 2026, Black Forest Labs (the German lab behind FLUX) shipped their first public video model, FLUX 3 Video, and the number they led with was 20 seconds of video with native audio in one pass.

The interesting part for makers is practical. One clip long enough for a full Reels beat, with sound baked in, if the early-access build holds up outside the demo reel.

This post is the video companion to our full FLUX 3 release guide. Image, editing, Action, and Dev live there. Here we stay on length, audio, and what you can actually plan around.

Quick answer: On July 23, 2026, Black Forest Labs put FLUX 3 Video into gated early access. One pass can make up to 20 seconds of video with native audio, including text-to-video and image-to-video. Pricing and open weights are not public yet. Longer films still need multiple clips, references, and an edit.

Note- I have not run FLUX 3 on TheFluxTrain yet. Everything below is from BFL's announcement and early press. Treat the feature list as claimed until you burn a few of your own prompts on it.

What did Black Forest Labs launch?

The short version: FLUX 3 is a whole family. Video is the door that is open right now.

You apply for gated early access. BFL approves. The product name you care about is FLUX 3 Video. Clips with optional native audio. Same launch also promised FLUX 3 Image in the following weeks, Action work with partners (including robotics), and FLUX 3 Dev open weights later with no date. If you want that roster in one table, jump to the FLUX 3 release guide.

Under the hood, BFL says they train one architecture across images, video, and audio. For your next client brief, that mainly means one generator that is supposed to keep picture and sound in the same pass.

What can FLUX 3 Video actually do?

BFL's list is long. The ones that matter for day-to-day work:

• text-to-video
• image-to-video from a starting frame or a reference still
• video-to-video that carries a subject (a character, a product) into a new scene
• continuation from existing video and audio
• keyframe-controlled transitions
• multilingual dialogue
• chaining clips into longer multi-shot sequences
• varied styles and aspect ratios, plus animated type

If I got a key tomorrow, I would start with audio sync and faces. Dialogue that lags the mouth, or a glass that "breaks" with no sound, kills trust faster than a soft background. BFL says facial expressions, multilingual speech, and event-matched sound are early strengths. That is exactly what I would score first.

One quiet gotcha: their published preference tests used 10-second, 720p clips. They have not stated the max resolution for a full 20-second generation. So when someone tweets "20s vs 1080p Tool X," you are not looking at the same yardstick yet.

What can you do with a 20-second AI clip?

Twenty seconds is the length of a lot of things people actually ship:

• a Reels or TikTok beat with room to hook and show the product
• a product hero move with voiceover in one generation
• one dialogue beat in a short film instead of three awkward cuts

Picture a skincare bottle on a bathroom shelf. The old AI path was a still, then a 4s orbit, then a 4s pour, then a 4s smile, and a prayer that the face still matched. The new pitch is one 15-20s pass with the pour sound and a line of VO in the same file. That is the promise. Whether it holds depends on access and how stubborn your brand look is.

For 30-60 second brand films, BFL talks about chaining clips with visual references so characters survive across scenes. That is useful. It is still production. Shot list, references, edit. You just stitch less.

When I ran an animation company years ago, before any of this AI video existed, we lived in After Effects libraries and reusable motion snippets so the next spot did not start from a blank timeline. Nobody asked "can the render go longer?" They asked "will Thursday's version match Monday's character?" Continuity ate the calendar. A 20-second max helps. It does not retire that problem.

How does 20 seconds compare to other AI video models?

Launch-week snapshot. Prices and access change weekly, so double-check before you budget.

BFL's early preference chart looked strong against Luma Ray 3.2 (93%) and Runway Gen-4.5 (77%), and basically tied Seedance 2.0 and Gemini Omni Flash (~52%). Those numbers are preliminary, from an early candidate, on 10-second clips. They are fun to read and a weak purchase decision on their own.

What I would watch instead is simpler: whether I can get a key, what a redo costs, whether the character holds for three scenes, and whether the audio survives a loud café ambience prompt. Length is the headline. Those answers decide if it enters a weekly workflow.

What is available now vs later?

So if a headline says "FLUX 3 is out," ask which door. Right now the honest answer is: video early access for people BFL lets in. APIs, private weights, and broader editing access are on the roadmap, not in your hands yet.

Does a longer clip replace a real video workflow?

A 20-second pass is a nicer building block than a 5-second one. You still need the rest of the process around it.

Start with a brief: who it is for, how long, which platform, one job for the clip. Lock looks for people and products when brand matters. If you chain scenes, write the shot order before you burn credits. Then edit for pacing and captions the way you would after any other shoot.

That full path is what we call vibe video: talk the story through with an agent, then run characters, scene clips, and a timeline in a studio. Models will keep stretching the max length. The people who ship every week are the ones who stopped starting from a blank prompt each Monday.

TheFluxTrain is that production side for us. When FLUX 3 access opens wider, we will try it in Flow Studio. We will say so when it is actually wired. Not before.

How should creators try FLUX 3 Video today?

1. Read the source: FLUX 3 announcement.
2. Apply for early access if video is real work for you this quarter, not just curiosity.
3. Freeze a tiny test kit before you get cute: one text prompt, one product still, one character reference, one dialogue line. Score face lock and audio sync. Count redos.
4. Anything over ~20 seconds gets a shot list so you know where chaining starts.
5. For image, editing, and Dev status, stay on the FLUX 3 release guide.

Already making short films or product spots with agents? Start from What is vibe video? and drop the new model into that structure when you get access. Swap the generator when something better shows up. Keep the shot list.

Frequently asked questions

Who launched the 20-second AI video model?

Black Forest Labs, based in Freiburg, Germany. FLUX 3 Video was announced on July 23, 2026 as part of the FLUX 3 multimodal family.

Is FLUX 3 Video free or publicly available?

Gated early access. You apply, BFL approves. No general public API or published price list yet.

Does every FLUX 3 Video clip include audio?

BFL describes native audio with the video, including dialogue and sounds tied to on-screen events. Check it on your own prompts before you promise it to a client.

Is 20 seconds the maximum finished film length?

Twenty seconds is the stated max for one generation. Longer pieces need multiple clips chained with references, then an edit.

How does FLUX 3 Video compare to Runway, Luma, or Veo on length?

On paper it is among the longer single-pass lengths in this class, with evals shown at 720p. Resolution, price, and access differ by vendor. Read current docs, not launch-week charts alone.

Can you use FLUX 3 Video for ads and Reels?

A 15-20 second clip matches a lot of short-ad and Reels lengths. Rights, brand safety, and whether you have a key still sit with you and BFL's terms. We do not ship FLUX 3 templates on TheFluxTrain yet.

Is FLUX 3 open source?

BFL has announced a future open-weight model, FLUX 3 Dev. License, size, and hardware needs are not out. "Open-weight" is the accurate phrase for now.

When will FLUX 3 be on TheFluxTrain?

Not announced. We try new video models as access opens. Watch the blog and changelog; this post is not a ship date.

Full family details (image, editing, Action, Dev) are in the FLUX 3 guide. Curious how longer clips might land in Flow Studio later? Leave a comment and I'll answer what I can.

Cheers!