FLUX 3 Max: Video Fast Enough to Iterate On

Most AI video work isn't creative work. It's waiting. You write a prompt, send it off, and then sit there for minutes while the render finishes, only to find the camera drifted left when you wanted it to push in. By the end of the afternoon you've seen a dozen takes and settled for the eighth because you ran out of patience, not because it was right.

Speed decides how many ideas you get to test before you commit.

That's the case for FLUX 3 Max, the newest video tier from Black Forest Labs, released on October 19, 2026. BFL calls it its fastest and highest quality video model, and it's live on each::labs today as Text to Video, Image to Video and Keyframes to Video. The pitch is simple: a ten-second clip in roughly twelve seconds. That changes how you prompt, not just how long you wait.

What FLUX 3 Max Actually Is

FLUX 3 Max sits on top of FLUX 3 Video, the video side of Black Forest Labs' FLUX 3 family. According to VentureBeat's launch coverage, it builds on FLUX 3 Video with additional training aimed at three things: following instructions more closely, producing better-looking output, and cutting generation time. BFL's own FLUX 3 Video page sums up the Max section in one line: a 10-second video in 12 seconds, which BFL says is 22 times faster than ByteDance's Seedance 2.5 at the same quality.

That's a vendor claim, and you should read it as one. VentureBeat's report puts the average at 12.45 seconds for a ten-second, 16:9 clip at 720p, measured by the launch team, not an outside lab. On quality, the same report points to the Artificial Analysis image-to-video leaderboard as checked on launch day: FLUX 3 Max at 1,138 Elo, level with Seedance 2.5 and two points behind MiniMax H3. Statistically, that's a tie at the top. Max plays in the same league as the best image-to-video models while finishing far sooner.

The clip itself stays familiar: five to twenty seconds in a single generation, native audio on by default, and aspect ratios from 21:9 ultrawide down to 9:16 vertical.

Speed Changes How You Prompt

When a render takes several minutes, you prompt defensively. You pack every instruction into one long paragraph, because each attempt eats minutes of your attention, and you hope the model reads all of it. When a render takes about as long as it takes to reread your prompt, you can work the way a director works with a camera operator: one note at a time.

Try a take. Watch it. Change the light. Try again. Generated video mostly lost that loop because feedback came too late. A model that answers in seconds keeps you inside the idea.

Here's a prompt short enough to edit one clause at a time:

A lighthouse keeper climbs a spiral iron staircase at dusk, lantern swinging in one hand. Slow upward tracking shot following him. Wind whistling through gaps in the stone, boots ringing on metal steps.

Run it, then swap "dusk" for "a storm at midnight" and run it again. Then change "slow upward tracking shot" to "static shot from the top of the stairs looking down". Three variations, three clear answers. More takes is not the goal. Faster answers to specific questions is.

Three Ways In: Text, Image, Keyframes

The three FLUX 3 Max models on each::labs share almost every parameter. What separates them is how much of the shot you decide before the model starts.

FLUX 3 Max Text to Video decides nothing in advance. You send a prompt, pick a duration and a ratio, and the model builds the whole frame. With aspect ratio left on auto, it defaults to 16:9. This is where you explore: test a concept, a mood, a camera move, before you've made any assets for it.

FLUX 3 Max Image to Video starts from a picture you supply as the first frame, with an optional end image as the last frame. Give it both and you're no longer asking for motion, you're asking for a path between two approved compositions. On auto, the output follows the ratio of your first image, so a vertical product shot stays vertical.

The ceramic teapot slowly rotates on the wooden table as steam curls from its spout. Soft morning window light, shallow depth of field, gentle clink of porcelain and a kettle hissing in the background.

That prompt describes how the image unfolds, not what's in it. Spend your words on motion, camera and sound.

FLUX 3 Max Keyframes to Video lets you pin images to specific moments. It's the most controlled of the three.

Pin the Story With Keyframes

The Keyframes model takes between one and ten images, and each one gets a frame index: the exact frame of the 24 fps output where that image should land. A keyframe at index 0 opens the clip. One at index 120 lands at the five-second mark. On a twenty-second clip, the highest index you can use is 480. Every index has to be unique, which keeps you from stacking two images on the same moment.

That turns a prompt into something closer to a storyboard. Say you're building a twelve-second reveal for a watch. Frame 0 shows the closed box. Frame 96, four seconds in, shows the lid open. Frame 240, ten seconds in, is the hero close-up. The prompt then only has to describe the movement between them:

The camera glides in as the velvet box lid lifts on its own, revealing the watch. A slow orbit around the dial as light sweeps across the polished steel. Quiet mechanical ticking, a soft orchestral swell at the close-up.

Our guide to pinning FLUX 3 Video keyframes to seconds covers the frame math, and it carries straight over to Max. Decide the beats first, then write the prompt.

FLUX 3 Max vs FLUX 3 Video: What Actually Changes

Here's the straight comparison, built from the request schemas of the each::labs models. Both tiers run five to twenty seconds, both generate audio by default, and both offer the same eight aspect ratios. The differences are in the controls around them. Standard FLUX 3 Text to Video accepts "auto" as a duration, letting the model fit the length to the content. Max wants an explicit number of seconds, defaulting to five. Standard FLUX 3 picks between an hd class (up to one megapixel per frame, around 1280x704) and an fhd class (up to two megapixels, around 1920x1088, finished by BFL's video upsampler). Max offers 720p and 1080p, defaulting to 720p.

Standard FLUX 3 has a draft mode: a quick hd preview that keeps a reference so you can re-render the keeper at full quality later. Max has no draft switch, which makes sense: when the full render is this fast, a preview step stops pulling its weight. In exchange, Max adds two controls the standard text model doesn't have: a reasoning level and a seed. The seed lets you hold a take steady while you change one other variable, which is the backbone of disciplined iteration.

The family stays wider on the standard side. FLUX 3 Video on each::labs also includes Video Continuation, which extends an existing clip from its final frame, plus a timestamped Image to Video model. Max covers text, image and keyframes. So the practical split looks like this: reach for Max when you're iterating fast or need its tuned instruction following, and keep standard FLUX 3 for continuation work and the auto duration that suits loose briefs. Our earlier post on how FLUX 3 thinks in scenes, not clips covers the multi-shot side that both tiers share.

Turn Reasoning Up or Down

The reasoning level is Max's most interesting control and the easiest one to ignore. It sets how much effort goes into working through your prompt before generation starts. It has two settings, low and high (the default), and the schema is blunt: higher reasoning means slower generation.

Treat it like a gear shift. When you're testing whether a concept works at all, drop to low and burn through variations. Once the idea is settled and the prompt carries more weight (several characters, a specific sequence of actions, a line of dialogue that has to land), switch back to high and let the model think. A complex prompt is exactly where closer instruction following pays off.

Two chess players in a smoky 1920s café. The older man taps the clock, the younger one hesitates, then slides his queen across the board. The older man leans back and says quietly, "Well played." Camera holds a medium two-shot, then cuts to a close-up of the board. Murmured conversation, a spoon stirring coffee.

An order of events, a cut and a line of speech: that's a high-reasoning prompt. Drifting fog over a lake is not.

Getting Better Results From FLUX 3 Max

Lock the seed when you iterate. If you change the prompt and the seed at the same time, you can't tell which change made the difference. Fix the seed, change one clause, compare.

Choose the ratio before you write. A crowded street scene written for 21:9 reads very differently at 9:16.

Write the sound into the prompt. Audio is on by default, so name the ambience, effects and dialogue. VentureBeat notes BFL highlights speech, physical behavior and sound effects as strengths, so ask for them.

Prototype at 720p, deliver at 1080p. Once motion and timing are right, rerun the same prompt and seed at 1080p for the version that ships.

Start short. Prove the motion in five seconds, then stretch to fifteen or twenty once you trust it:

Vertical 9:16. A street dancer spins on a wet rooftop at night, neon signs reflecting in the puddles beneath her. Low angle, camera slowly circling. Bass-heavy music from a distant speaker, rain pattering on metal.

Running FLUX 3 Max on each::labs

All three Max models run through the same each::labs API as the rest of the catalog, so switching between them is a change of model name and input shape. A keyframes request input looks like this:

{
  "prompt": "The camera glides in as the box lid lifts, revealing the watch. Slow orbit around the dial.",
  "keyframes": [
    { "image_url": "https://example.com/box-closed.png", "frame_index": 0 },
    { "image_url": "https://example.com/box-open.png", "frame_index": 96 },
    { "image_url": "https://example.com/watch-closeup.png", "frame_index": 240 }
  ],
  "duration": 12,
  "aspect_ratio": "16:9",
  "resolution": "720p",
  "reasoning_level": "high",
  "generate_audio": true,
  "seed": 42
}

Swap to the Image to Video model and replace the keyframes array with an image_url, plus an optional end_image_url. Swap to Text to Video and drop the images entirely. Everything else stays the same, so you can explore in text, lock the opening frame, then storyboard the final cut with keyframes on one integration. If you're comparing Max against other partners' models for the same brief, the image-to-video models on each::labs sit behind that same call, and our ranking of the best AI video generation models in 2026 is a good map of when to pick which.

Wrapping Up

Black Forest Labs didn't add a new trick with FLUX 3 Max. It made the existing ones fast enough to use the way you'd like to: twenty-second clips, native audio, first and last frames, ten pinned keyframes and a reasoning dial, answering in seconds rather than minutes. The quality holds level with the top image-to-video models. The speed is the reason to try it.

The best take is rarely the first one. It's the one you got to because the previous five came back quickly. Open FLUX 3 Max on each::labs and see how many ideas you can test before your coffee cools.

Frequently Asked Questions

How fast is FLUX 3 Max compared with FLUX 3 Video?

Black Forest Labs says a ten-second FLUX 3 Max clip renders in about twelve seconds, and VentureBeat's launch report gives an average of 12.45 seconds at 720p. BFL's model page doesn't give a matching figure for standard FLUX 3 Video, so the cleaner reference point is BFL's own comparison: 22 times faster than Seedance 2.5 at the same quality. Time a few renders of your own.

What does the reasoning level do in FLUX 3 Max?

It controls how much effort the model spends working through your prompt before it generates. High is the default and suits prompts with several actions, characters or dialogue. Low is quicker, which makes it the better setting for early exploration when you're testing many variations of a simple idea.

Can FLUX 3 Max use keyframes?

Yes, through the dedicated Keyframes to Video model. You can pin up to ten images to exact frame positions in a 24 fps clip, each with a unique index, up to the last frame of the clip. For simpler jobs, Image to Video takes a first frame and an optional last frame.

Should I use FLUX 3 Max or standard FLUX 3 Video?

Max suits fast iteration and prompts that need tighter instruction following, and it adds a seed and a reasoning control. Standard FLUX 3 keeps auto duration, draft previews and video continuation. Many projects will use both, and on each::labs they share one API.