FLUX 3 Max Explained: Top-Ranked Image-to-Video Quality, 22× Faster (October 2026)

October 20, 2026 · 7 min read

FLUX 3 Max is a video generation model released on October 19, 2026 by fal, post-trained by fal Research in collaboration with Black Forest Labs on BFL's FLUX 3 foundation model — and its claim to attention is simple: image-to-video quality tied with the best models on the leaderboard, generated 10 to 22 times faster than the flagships it ties with. It's the second model of its kind from fal, after H3 Max did the same trick to MiniMax's H3 in August (covered in our H3 Max explainer): take a strong foundation model, keep training it for fidelity, motion coherence and prompt adherence, and co-design the serving path so the quality arrives at iteration speed.

What FLUX 3 Max actually is

The launch pairs two teams with distinct roles. Black Forest Labs built the FLUX 3 foundation model — the multimodal backbone that generates images, video with native audio, and more (the image side is covered in our FLUX 3 Image guide). fal Research then continued training that model — post-training, reinforcement learning and inference optimisation — and tuned how it executes on fal's own inference engine. BFL CEO Robin Rombach framed it as the collaboration FLUX was designed for: a foundation model meant to be built on.

At launch, FLUX 3 Max serves text-to-video and image-to-video at 720p through fal's API, with keyframes-to-video announced alongside. Content generated through the API can be used commercially under fal's terms. fal is explicit that this is the first iteration — further post-training for quality and speed is already underway.

The numbers behind the launch

Two claims matter, and both are worth reading carefully. First, quality: fal reports FLUX 3 Max at an 1,139 Elo on Artificial Analysis's preliminary AA-Video-I2V v1.1 ratings (4,868 votes) — a 50-point jump over the original FLUX 3 (1,089) and level with Seedance 2.5. Artificial Analysis's own board lists it at 1,138 ± 15, second behind MiniMax H3 at 1,140 — within the error bars of first place either way you read it.

Second, speed — the actual headline. fal's benchmark renders a 10-second 720p clip in about 11 to 12.5 seconds. Against the same 10-second 16:9 test in fal's published comparison:

That's the "22× faster than Seedance 2.5 at the same quality" banner on BFL's homepage, and roughly 10–11× faster than the FLUX 3 it was trained from. The one caveat fal itself notes: its own H3 Max is still faster in this comparison — so the honest framing is flagship-quality at near-draft speed, not the outright speed record. And the standard caveat applies too: these are preliminary, vendor-published benchmark readings, so treat exact figures as directional until independent testing settles in.

What it costs

Launch pricing on fal is $0.0425 per second of 720p output, billed per second with no subscription — a 75% discount off the $0.17-per-second list price, running until November 7, 2026. A 10-second clip at the launch rate costs about $0.43. For context inside Hermosso, the original FLUX 3 is already in the video picker — 850 credits for a 5-second 720p clip with audio on plans from $9/month — so the FLUX line is one engine you can benchmark today against Kling 3, Seedance 2.5, Hailuo H3 Max and the rest of the 18 video engines on a single credit balance.

Why "post-trained for speed" is the story of this model cycle

Twice now in two months, the biggest practical jump in AI video hasn't come from a new foundation model but from an inference company re-training an existing one. H3 Max took MiniMax's open H3 weights to #1 at faster-than-real-time rendering; FLUX 3 Max takes BFL's foundation to tied-#1 quality at a fraction of the flagship render time. The pattern matters for anyone planning production work: the bottleneck in AI video was never peak quality — the top models have been within a few Elo points of each other for months — it's the cost and latency of iteration. A 10-second clip in 12 seconds changes how you work: you can afford to try ten prompts and keep the best, where a four-minute render forces you to commit early. That is the same philosophy behind Hermosso's engine picker — run the same input through several models on one balance and keep the winner — pushed down into the model layer itself.

What you can do on Hermosso today

An honest status note first: FLUX 3 Max is days old and not yet in Hermosso's studio picker as of writing — new engines are added as they stabilise, the same way Hailuo H3 Max and Nano Banana 2.1 landed within days of their launches. What's already there:

• FLUX 3 — the foundation model FLUX 3 Max was trained from — in Photo to Video and Text to Video: 5-second 720p clips with native audio at 850 credits.
• Hailuo H3 Max — fal's previous post-trained model and still the speed benchmark — at 500 credits per clip with sound, the cheapest flagship-tier engine in the picker.
• The rest of the field — Kling 3 and 3 Turbo, Seedance 2.5, Veo 3.1, Runway Gen-4.5, Wan 3.0 and more — so the same still or prompt can go through three engines in minutes. New accounts get 400 free credits, and what every engine costs in credits is public.

When FLUX 3 Max is NOT the right choice

The honest limits, because every launch page skips them. If you need long, directed scenes, this isn't the tool — Kling 4.0, released four days earlier, does single takes up to 30 seconds with 10 keyframes and 15 references; FLUX 3 Max at launch is a short-clip, 720p model. If you need resolution above 720p — 1080p or 4K delivery — wait for higher-res endpoints or use a model that ships them today. If your work is recasting real footage — putting a specific person into an existing clip — that's a different technique entirely, and it's what Motion Transfer is for. And if the launch discount matters to your budget, note the date: the $0.0425 rate ends November 7, after which the math changes at list price.

What to watch next

Three things. Whether the Artificial Analysis placement holds as more votes accumulate — preliminary boards move. Whether fal ships the announced keyframes-to-video endpoint and resolutions above 720p, which would open the model to longer, directed work. And whether fal's continued post-training keeps the quality curve rising at the same speed — the team has said publicly this is iteration one. We update this page when any of those land.