FLUX 3 Video: 20-Second Clips With Native Audio

What FLUX 3 Video is and when it launched

FLUX 3 Video is Black Forest Labs' video generator with sound built in, and it reached general availability on August 4, 2026. It creates video from text, images or an existing clip, and the same model produces the dialogue and effects alongside the picture. Access is through the BFL API or a select partner.

FLUX 3 Video at a glance

Made by

Black Forest Labs (BFL)

Generally available

August 4, 2026

You send

A text prompt, 1 to 10 images, or a clip to continue

You get

5 to 20 seconds of video at 24 fps with generated audio

Where

BFL API and select partners, or the browser Playground

Status

The docs call FLUX 3 a preview model

The August 4, 2026 general availability

On August 4, 2026, Black Forest Labs opened an initial version of FLUX 3 Video, covering text and image input, to everyone. The broader FLUX 3 model had been announced in July 2026 with an early access phase, so August is when the video part left early access.

According to the August post, the video capabilities had advanced since that first announcement. When dating the FLUX 3 release, use August 4 for video and July for the announcement of the model family.

How FLUX 3 Video relates to the wider FLUX 3 model

FLUX 3 is the parent model, and FLUX 3 Video is the part that generates video. Black Forest Labs describes FLUX 3 as its frontier multimodal model, built to generate and predict video, audio, images and actions. August marked the first time its video generation was open for general use.

The other parts have not followed yet. FLUX 3 Image, the image generator, is still on the roadmap, and the action model is a separate open-weight release. The docs call FLUX 3 a preview model, so the details here are as of September 30, 2026.

Where you can use FLUX 3 Video today

There are three ways in: the BFL API, select partners, or a browser Playground that needs no setup. The API requires an account with credits, and the blog mentions "select partners" without naming them.

Editing an existing clip is handled by a separate product. FLUX Video Edit [fast] has its own endpoint, /v1/flux-tools/video-edit-v1, and is not part of the flux-3-video endpoint.

What you can make with FLUX 3 Video

FLUX 3 Video builds clips from a text prompt, from images pinned as keyframes, or from an existing clip, and it generates the sound together with the frames. The API exposes these as three modes, t2v, i2v and v2v. One clip can also contain several shots.

The three modes of the flux-3-video endpoint

t2v (text to video)

Starts from your prompt alone

i2v (image to video)

Starts from your images, which become frames of the clip itself

v2v (video continuation)

Starts from a clip you already have and keeps it going

Text to video and image to video with keyframes

Text to video only needs a prompt, whereas image to video begins with one to ten images pinned as keyframes. A single image becomes the exact first frame. With two images, the start and end are fixed and the model works out the motion in between.

Each image can also be pinned to a specific second, so up to ten of them act as a storyboard. The prompt then describes what happens around those frames.

How video continuation carries a clip forward

Hand the model footage you already have and it picks up where that footage ends, carrying motion, camera behavior, dialogue and audio across the join. You supply the clip in a start_video field and describe what happens next, and generation continues from its final frames.

The blog describes using up to four seconds of existing video and audio as input. That suits footage you have already shot or generated, because the new section inherits its momentum and sound instead of starting cold from a prompt.

Multiple shots, dialogue and sound in one clip

One generation can cut between camera angles, and the model writes the speech, effects and background noise together with the picture. According to the docs, characters, look and continuity stay consistent across hard cuts, with one audio bed running through. Their multi-shot example writes out each shot in the prompt and describes the cuts in words.

Audio is on by default, and setting generate_audio to false gives a silent clip. For speech, put the line in quotes in the prompt and the character says it. The blog adds that output need not look cinematic and that text can appear inside the scene.

Which languages the lip-syncing covers

Black Forest Labs names 13 for spoken lines, then adds "and more". They include Chinese, Japanese, Hindi, Punjabi, Indonesian and Turkish, along with Spanish, French, German, Portuguese, Italian, Russian and English in several dialects.

The company calls the lip-syncing precise. That is its own claim, and the post offers no quality figures per language, so test your language before relying on it.

Clip length, resolution and frame rate

One request returns 5 to 20 seconds of video at 24 fps (5 to 15 for continuation), in a resolution class from hd up to uhd, which reaches 3840 x 2176 at 16:9. The blog and the docs describe the output differently: the blog mentions only HD and Full HD, while the docs go up to 4K.

How long a FLUX 3 Video clip can run

Text to video and image to video produce 5 to 20 seconds, while video continuation produces 5 to 15. The 20-second headline figure therefore covers only the first two modes.

The duration field accepts whole seconds from 5 to the mode's maximum, or auto, the default. On auto, the model sets the length to suit the content.

HD, Full HD, QHD and UHD output classes

There are four output classes, hd, fhd, qhd and uhd, and hd is the default. The blog mentions only HD (720p) and Full HD (1080p), the latter reached by upscaling, while the docs add qhd and uhd, with uhd reaching 3840 x 2176 at 16:9.

Only hd is the direct output of generation. The other three take that same generation and finish it with the video upsampler on a bigger canvas, so the shot is the same in all four and only the delivered size changes. Delivered sizes are not always standard: uhd at 16:9 comes out at 3840 x 2176 rather than 3840 x 2160, so crop if you need an exact standard size.

Aspect ratios available and the fixed 24 fps

All modes output at 24 fps, and eight aspect ratios are available: 16:9, 9:16, 1:1, 4:3, 3:4, 2:1, 21:9 and 9:21. The aspect_ratio field defaults to auto, which adapts the frame to the content.

The tallest options, 9:16 and 9:21, fit phone-first formats. No field sets the frame rate, so projects needing a different rate must convert afterward.

FLUX 3 Video pricing per second of output

FLUX 3 charges per second of output at a rate set by the mode and resolution class, from $0.06 for a draft to $0.95 for uhd continuation. A clip costs its length multiplied by the rate for your mode and class. All rates are in US dollars.

Rates for text to video and image to video

Text to video and image to video use the same price list: $0.17 a second at hd, $0.29 at fhd, $0.40 at qhd and $0.80 at uhd. Drafts in either mode cost $0.06 a second, and since drafts always render at hd, there is no draft price for larger classes.

Larger classes are billed at the class delivered. Because fhd, qhd and uhd are the same generation upscaled, paying more buys a bigger canvas rather than a different shot.

Why video continuation costs more per second

Video continuation is pricier per second, at $0.41 to $0.95 for a full render and $0.12 for a draft, though the gap narrows at higher resolutions. Compared with text to video, it costs 2.4 times as much at hd, 1.8 times at fhd, 1.6 times at qhd and 1.2 times at uhd. Drafts cost twice as much.

The docs give no reason for the premium, so take it as a price rather than a cost breakdown. Continuation is also capped at 15 seconds, so it costs more per second and runs shorter than a new 20-second generation.

Worked example: what a 20-second clip costs

A 20-second text-to-video clip comes to $3.40 at hd, $5.80 at fhd and $16.00 at uhd, or $1.20 as a draft. Each figure is length times rate. Continuation maxes out at 15 seconds, so its examples are shorter.

• Text to video, 20 seconds at hd: $0.17 × 20 = $3.40
• Text to video, 20 seconds at fhd: $0.29 × 20 = $5.80
• Text to video, 20 seconds at uhd: $0.80 × 20 = $16.00
• Text to video, 20 seconds as a draft: $0.06 × 20 = $1.20
• Video continuation, 15 seconds at hd: $0.41 × 15 = $6.15
• Video continuation, 15 seconds at fhd: $0.53 × 15 = $7.95

How to test ideas cheaply with Draft mode

A quick hd preview costs roughly a third of a full render, so full price applies only to the shot you keep. Send draft: true for the preview, then return the approved one with mode: "draft_enhance" to finish it. The savings are real but get smaller with each extra attempt at the same shot.

What a draft costs against a full render

Drafts cost $0.06 a second for text to video and image to video and $0.12 for continuation, compared with $0.17 and $0.41 for a full hd render. On a 20-second text-to-video clip, that is $1.20 for the preview versus $3.40 for the render.

The number of attempts matters, though. Two hd drafts add up to $0.12 a second, still below one full render at $0.17, but a third pushes the total to $0.18. From the third draft of a shot, you have spent more than one direct render would cost, and the draft you keep still needs a full-price render on top.

How draft_enhance keeps the shot you picked

Keep the draft_cache bundle that the preview returns, and send it back to get the same generation at full quality. That second request uses mode: "draft_enhance", and the docs state that nothing is re-interpreted and that the shot and seed stay the same. The blog says subjects, composition and motion match the approved version.

# 1. Draft: a fast hd preview plus a draft_cache bundle in the result
curl -s -X POST https://api.bfl.ai/v1/flux-3-video \
  -H "x-key: $BFL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"mode": "t2v", "prompt": "a fox running through dawn mist", "draft": true}'

# 2. Enhance the one you picked, sending its bundle back
curl -s -X POST https://api.bfl.ai/v1/flux-3-video \
  -H "x-key: $BFL_API_KEY" \
  -H "Content-Type: application/json" \
  -d "{\"mode\": \"draft_enhance\", \"draft_cache\": \"$(base64 -i draft_cache.bin)\"}"

Drafts have to be hd, and a draft: true request at any other resolution is rejected. The draft_enhance step accepts all four classes and defaults to fhd, so you choose the final size when you commit rather than when you preview.

Why a fresh full render can look different

Each submission is its own generation, so resending the same prompt can give a different take on the shot from the draft you liked. To show this, the docs run one prompt down three paths: the draft, the enhanced draft and a direct full render.

That is the reason for the enhance step. Liking a draft and then resubmitting the prompt as a normal request starts a new generation at full price, with no guarantee it matches.

How to call the FLUX 3 Video API

Calls are asynchronous: you send a request to api.bfl.ai/v1/flux-3-video, poll the returned address until the job is Ready, and download the video from the result. Authentication uses an x-key header holding your API key, and every request specifies a mode.

The submit, poll and download round trip

The first response carries an id and a polling_url, and once that URL reports status as Ready, result.sample holds a signed .mp4 URL. Errors and moderation stops arrive as the statuses Error, Request Moderated and Content Moderated, and the loop below handles each one.

import os, time, requests

key = os.environ["BFL_API_KEY"]

submit = requests.post(
    "https://api.bfl.ai/v1/flux-3-video",
    headers={"x-key": key, "Content-Type": "application/json"},
    json={"mode": "t2v", "prompt": "a fox running through dawn mist", "generate_audio": True},
).json()

while True:
    time.sleep(2)
    result = requests.get(submit["polling_url"], headers={"x-key": key}).json()
    if result["status"] == "Ready":
        print(result["result"]["sample"])
        break
    if result["status"] in ("Error", "Request Moderated", "Content Moderated"):
        raise RuntimeError(result["status"])

According to the docs, only mode and prompt are always required, even though their own draft_enhance example sends just mode and draft_cache. i2v needs keyframes, v2v needs start_video and draft_enhance needs draft_cache. The optional resolution, duration, aspect_ratio, generate_audio, safety_tolerance, draft and version fields have defaults, and version defaults to latest, which follows the current release.

Requests and responses are JSON, and their shape varies by mode. Check a sample response from each mode you plan to use before writing the parser.

How long result URLs stay valid

Result URLs are signed, and the docs say they expire about 2 hours after the job finishes. Download the video as soon as the status turns Ready.

The 2-hour window is not a safety margin. If a job finishes overnight and nothing is polling, the URL will have expired by morning, so fetch the file inside the polling loop.

Where to try it without writing code

The Playground runs FLUX 3 in a browser with no setup. It is the fastest way to see whether a prompt style works for you before writing API code. It costs the same, though, since the pricing page lists one price for both the Playground and the API.

Using the API requires an account with credits. The quickstart page covers creating the account, adding credits and making the first call.

Are the FLUX 3 weights open?

FLUX 3 Video cannot be downloaded; it runs only through the BFL API and select partners, and FLUX 3 Action is the one FLUX 3 model you can download today. FLUX 3 Dev, the open-weight variant meant for content creation, is still on the roadmap.

What is open today: FLUX 3 Action

FLUX 3 Action is an open-weight 7B world action model that predicts robot actions rather than producing video for an audience. From camera images, the current state and an instruction, it outputs a chunk of actions plus about two seconds of predicted video.

It can drive a robot arm through LeRobot or be fine-tuned for your own robot, simulator or game. Despite sharing the FLUX 3 name, it offers no way to run FLUX 3 Video on your own hardware.

What is planned: FLUX 3 Dev and FLUX 3 Image

FLUX 3 Dev (an open-weight variant) and FLUX 3 Image (image generation and editing) are on the roadmap and not yet released. The August post adds that upcoming releases will bring finer control over video and generation from mixed image, video and audio references.

The July announcement from Black Forest Labs describes the open-weight release as access to a multimodal backbone for video, audio, image and action prediction. Neither post gives a date for it.

What this means for running it locally

FLUX 3 Video cannot run on your own hardware today, so every clip is generated on servers operated by Black Forest Labs or its partners and billed per second through the API. If footage or prompts must stay in-house, the model is off the table until an open variant ships.

Neither apparent workaround helps. FLUX 3 Action predicts robot actions instead of making clips, and FLUX 3 Dev has no release date, so plan around it only after it ships.

How good is FLUX 3 Video compared with rivals?

Black Forest Labs claims FLUX 3 Video leads at text to video and ties Seedance 2.0 at image to video, but that comes from its own evaluation, not an independent test. The claim is light on detail, and the post names just one rival.

What BFL's internal evaluation claims

Black Forest Labs says its own tests put FLUX 3 ahead of existing state-of-the-art models at text to video by a solid margin. For image to video, the company says it ties Seedance 2.0 and beats every other model. It adds that human raters preferred it for both.

No scores are given in the post. Nor does it say who the raters were or which prompts they saw.

Why the Seedance 2.0 tie needs a caveat

Matching one named model says little about the unnamed ones, and a test run by the model's own maker is not a third-party benchmark. Seedance 2.0 is the only model the post names, as the one it ties. The models it claims to beat are not listed.

As a result, there is no way to check which models were compared, how many comparisons were run or how prompts were selected. Treat the claim as a reason to shortlist FLUX 3 Video, then test your own prompts against the rivals that matter to you.

How BFL says it tested for misuse risks

Black Forest Labs says it brought in an outside partner, Cinder, to check FLUX 3 Video for abuse before release. The aim was to confirm that safeguards hold in every modality the model supports, with non-consensual intimate imagery (NCII) and child sexual abuse material (CSAM) named among the risks.

The company describes several layers of mitigation before, during and after release. Because the post shares no results, it describes a process rather than a measured outcome.

What FLUX 3 Video still lacks today

FLUX 3 still carries a preview label, Omni Reference is not yet available, the video model has no weights, and FLUX 3 Image and FLUX 3 Dev are unreleased. Omni Reference is the planned mode for generating from image and video references. For now, inputs are text, images and a single existing clip.

Under the current docs, a request is limited to 20 seconds, or 15 for continuation. Since the version field follows the latest release by default, the same prompt may give different results after Black Forest Labs updates the model.