Wan 3 Is Coming: What We Know, and What We’re Doing About It

Alibaba hasn’t announced Wan 3.0. As of today, Wan 2.7 is still the latest generally available model in the series, and nothing called "Wan 3.0" appears in Alibaba Cloud Model Studio or the DashScope API. But on July 31, the AI tool platform PixelDojo posted that it had received early access to a next-generation Wan model and published demos of single-generation 30-second videos with sound. That, plus a July wave of Apache 2.0 ecosystem releases from the Wan team (WanSong for music, Wan-Dancer for music-to-dance, Wan-Streamer for real-time interaction), makes it look like a flagship release is close.

We’ve put up a Wan 3.0 page marked coming soon, and we’ll have it live behind the same endpoint and API key as the rest of the catalog when it ships. This post is what we actually know, separated from what’s being made up about it.

Why the Wan Series Is Worth Paying Attention To

Wan is Alibaba Tongyi Lab’s video model family, and it moves faster than almost anything else in the space: roughly six major releases in about a year.

The moment that mattered was February 2025, when Alibaba open-sourced the Wan 2.1 series under Apache 2.0, including a 1.3B version that ran on a single consumer GPU. Two million-plus downloads later, an entire ecosystem of ComfyUI workflows and LoRA fine-tunes grew around it. Wan 2.2 followed in July 2025 with a Mixture-of-Experts architecture: 27B total parameters, but only 14B active per step, splitting the work between a high-noise expert that handles scene layout and motion and a low-noise expert that handles detail and consistency.

From 2.5 onward the series added native audio, then stronger reference-to-video in 2.6, and landed on Wan 2.7 in April 2026: a four-mode suite covering text-to-video, image-to-video, reference-to-video, and instruction-based editing, at up to 15 seconds of 1080p with synchronized audio. Somewhere in there an anonymous model called "HappyHorse 1.0" topped the Video Arena leaderboards in both text-to-video and image-to-video Elo, and turned out to be from the same family.

What the Early Demos Actually Show

Everything below comes from third-party early-access testing in late July 2026, not from Alibaba. Treat it as a strong signal about direction and a weak signal about final numbers.

Clips up to 30 seconds in a single pass. This is the headline. Wan 2.7 caps out around 15 seconds; the demos show 2 to 30 seconds generated in one shot, no stitching. Going from 15 to 30 seconds is not just "twice as long" — temporal coherence decays fast in diffusion video, and most models start drifting past 10 seconds. Getting a stable 30 seconds means something changed architecturally.

Intelligent duration. Leave the length unset and the model picks it. In one test, "a paper boat washing down a gutter" came back as a 20-second clip with no duration specified anywhere in the request.

Audio generated with the picture, not after it. Dialogue, sound effects, room tone, and music come out of the same generation pass, embedded in the output file. The tell that this is architectural rather than a bolted-on post step: turning sound off reportedly costs exactly what leaving it on costs. One demo ran a 30-second kitchen scene with four lines of dialogue matched to lip movement.

Mixed multimodal references. Roughly 10 images, 5 video clips, and 5 audio clips in a single request — up from five references in Wan 2.7. Character from one input, environment from another, voice from a third, combined however you describe in text.

Consistency that survives the full take. In a 30-second rescue scene, two characters’ clothing, faces, and spatial relationships held throughout. That’s the difference between footage you can cut into a story and footage where someone quietly changes face mid-shot.

Adaptive aspect ratio. Prompt for a "vertical social ad" and you get 720x1280; prompt for a "square product loop" and you get 960x960. Neither request specified dimensions.

What Isn’t True

A cluster of third-party sites has been publishing Wan 3.0 specs that don’t match any demo or any official document: native 4K, 60 fps, 60-second clips, and a "Neural Physics Engine." None of that holds up.

The demos are 1080p. Physical realism does look better — pouring honey, drifting paper, cloth movement — but that reads as better training data and temporal modeling, not a separate physics module; integrating a real physics engine into a diffusion model is still an open research problem. Some sites also state release dates and open-weight licensing as settled fact. They aren’t. There’s a community rumor pointing at early August and nothing official behind it.

If you’re planning a production around this, use the demos and Alibaba’s own docs as your baseline, not the SEO pages.

How It Stacks Up Against Seedance 2.5 and Kling 3.0

On the specs we can compare today:

Wan 3.0 and Seedance 2.5 tie on the thing that changes workflows most — a full narrative arc in one generation. Seedance is clearly ahead on reference volume, though its resolution actually tops out at 720p (ByteDance's original 4K announcement didn't ship). Kling 3.0 is shorter per clip but packs up to six director-planned shots into that window and has the most mature multilingual dialogue support.

Where Wan could separate itself is the series’ open-source DNA. Wan 2.1 and 2.2 shipped Apache 2.0 weights; 2.5 and 2.6 didn’t. If 3.0 returns to open weights, it becomes the obvious choice for anyone who needs to self-host, fine-tune, or keep generation on their own infrastructure.

What This Means If You’re Building on Video

The practical shift is that 30 seconds plus native audio moves a model from "footage supplier" to "short film producer." The usual pipeline — generate 5 to 10 second clips, stitch them, dub them, cut them — collapses into a single call for a whole category of work: social spots that fill a full TikTok or Reels slot, 15-30 second ad first drafts, dialogue scenes, narrated product demos.

It also reinforces something we see across our users: nobody ships on one model. Teams reach for Seedance when they need heavy reference control and a 30-second single pass, Kling for multilingual dialogue and tight multi-shot pacing, Flux 3, which landed here this week, for flat per-clip pricing, and they’ll reach for Wan 3.0 when they need a long single-take narrative or a path to self-hosting. The interesting question isn’t which model wins, it’s how quickly you can swap between them.

Being Ready on Day One

That’s the part we can do something about now. Every model on Apiframe sits behind the same POST endpoint, the same API key, the same async job pattern, and the same prepaid credits. Adding a new model to your product means changing one string.

Build against Wan 2.7 today:

curl -X POST https://api.apiframe.ai/v2/videos/generate \
  -H "X-API-Key: afk_your_api_key_here" \
  -H "Content-Type: application/json" \
  -d '{
    "prompt": "a paper boat washing down a rain-filled gutter, low angle, afternoon light, ambient street sound",
    "model": "wan-2.7"
  }'

You get a job ID back immediately, poll /v2/jobs/:id or set a webhookUrl and let us call you when the clip is ready. When Wan 3.0 launches, "model": "wan-2.7" becomes "model": "wan-3.0" and everything else in your integration stays exactly where it is. No new account, no new SDK, no separate Alibaba billing relationship, and no migration project.

Wan 3.0 is billed at 26 credits ($0.26) per second of output, any resolution, on the same prepaid credits as everything else, and it’s already listed on apiframe.ai/pricing.

Follow Along

The reliable places to watch are the Wan-Video GitHub org, the Wan-AI models on Hugging Face, and Alibaba Cloud Model Studio’s release notes. We’re tracking all three, and Wan 3.0 goes live on Apiframe as soon as there’s an API to point at.

Get an API key and start on Wan 2.7 now, or watch the Wan 3.0 page for the launch. If you want to argue about whether it’ll beat Seedance 2.5, our Discord is already having that conversation.