Wan AI: Alibaba's Wan 3.0 video model

What is Wan AI?

Wan is Alibaba's family of AI video generation models, also known as Tongyi Wanxiang. It turns a scene you describe, or a photo you upload, into a short video, and since version 2.5 it generates sound together with the picture.

The family has two sides. Versions 2.1 and 2.2 were released as open source under the Apache-2.0 license, so you can download the model files from GitHub and Hugging Face and run them on your own computer. The weights for 2.5, 2.6, 2.7, and 3.0 weren't released: those versions are available through Alibaba Cloud's API, Alibaba's own site wan.video, and API providers.

The newest version is Wan 3.0. According to Alizila, Alibaba's news site, it opened in beta on Aug 6, 2026, and according to the Alibaba Cloud announcement it became generally available on Aug 24, 2026. Alibaba says it makes 30 fps video in 480p, 720p, and 1080p, from 2 to 30 seconds long, with dialogue, music, and sound effects.

Wan versions: from 2.1 to 3.0

The table is based on Alibaba's GitHub repositories and Alibaba Cloud announcements, checked in September 2026. Durations and resolutions are what the vendor offers on its own platform; the options available on Klakix are in the next section.

• Sources: 2.1 repository, 2.2 repository, 2.5 announcement, 2.6 announcement, 2.7 announcement, 3.0 general availability.
• The official 3.0 repository on GitHub holds only a README and an Apache-2.0 license file. The README mentions "open-source," but there are no model files to download.
• In 2026 Alibaba kept releasing open models separately: Alibaba's Hugging Face account lists Wan-Dancer-14B, which makes dance videos from music (July 2026), and Wan-Animate-2 (Aug 7, 2026), both under Apache-2.0.

What is Wan 2.5?

Wan 2.5 is the version Alibaba announced as a Preview at its Apsara 2025 conference on Sep 24, 2025. According to Alibaba Cloud's post, it was the first in the family to generate sound with the picture, and it raised the clip length from 5 to 10 seconds. According to the Model Studio documentation, it makes 5 or 10 second videos in 480p, 720p, and 1080p.

Its model files weren't published on Hugging Face; it's only available through the API and Alibaba's own platforms. So unlike 2.2, you can't run it on your own computer.

Wan 2.5 isn't offered on Klakix; we offer the newest version, 3.0. It keeps the built-in sound that 2.5 introduced and stretches a single generation to 30 seconds.

Using Wan 3.0 on Klakix

You'll find the model in both Text to video and Image to video. Describe the scene in plain words; with Enhance prompt on, your text becomes the detailed description the model understands, keeping your intent and details. Prompts can be up to 2,000 characters.

Saying how the camera moves and what you want to hear helps. One of our Wan samples was made from this prompt: “Sunset on the Galata Bridge: anglers casting their lines, the New Mosque and ferries behind, the camera moving slowly along the bridge. Seagulls and waves.”

• Duration: 5, 10, 15, 20, or 30 seconds; the default is 10. In-between lengths such as 2 seconds, and 25 seconds, aren't offered here.
• Resolution: 480p, 720p, or 1080p; the default is 720p. You can go up to 30 seconds in Full HD too.
• Aspect ratio: 16:9, 9:16, 1:1, 4:3, 3:4. There's no 21:9 widescreen option.
• Sound: on by default, and you can turn it off. Silent clips don't cost less; the price depends only on duration and resolution.
• In image to video the start photo is required and the end frame is optional: give two frames and the model builds the transition between them. Match photo is the default aspect ratio, so the video takes its shape from your photo without cropping. A prompt is optional in image to video.
• The real-face rule from the Seedance family doesn't apply to this model. The provider's general safety filter still does, and if a generation is blocked, your credits are refunded automatically.
• Generation runs in the background, and we estimate about 3 minutes. You can close the page; the finished video waits in your Library as an MP4.

How many credits is Wan 3.0?

The price depends only on duration and resolution, and it's the same for text and image to video, with or without sound. The numbers in the table come from our model catalog; the exact cost shows above the Generate button before you start.

• The default setting (720p, 10 seconds) costs 100 credits, about $3.63–$6.67 depending on the pack you buy.
• The longest, sharpest setting (1080p, 30 seconds) costs 600 credits. Trying the idea at 480p for 5 seconds first (25 credits) and extending once you like it saves credits.
• The smallest pack is Mini: 75 credits for $5. See all packs on the pricing page.

Is Wan AI free?

Alibaba's own site, wan.video, has a free plan. According to its pricing page, the free plan gives credits for a daily check-in and lets you generate one video and one image at a time; watermark-free downloads are listed as a Pro feature. Pro costs $5 a month billed yearly (list price $10) and gives 300 credits a month, which the page says covers up to 60 videos, and unlocks 1080p and 10 to 30 second videos. Premium costs $20 a month billed yearly. API usage is billed separately. We checked these prices in September 2026, and they may change.

To be upfront, there are cheaper or free ways to use these models: the free plan and low-cost Pro plan on Alibaba's own site, and, if you have a capable graphics card, the open-source 2.1 and 2.2 versions. What Klakix adds isn't the model itself: it's many video models in one account, pay-as-you-go credits, and no required subscription.

Klakix doesn't offer free video: we pay the AI provider for every generation. The 20 free credits you get after verifying your email work for voiceovers only. Wan 3.0 unlocks after your first credit purchase, and its cheapest setting (480p, 5 seconds) costs 25 credits. Before you buy, you can watch unedited Wan videos we made on the examples page.

Running the open-source versions on your own computer

The model files for 2.1 and 2.2 are open under Apache-2.0; code and setup steps are in the READMEs of the 2.1 and 2.2 repositories. You'll need a powerful graphics card (GPU). According to the 2.1 README, a 5-second 480p video takes about 4 minutes on an RTX 4090, and 2.2's 5B-parameter TI2V model runs at 720p and 24 fps on a 4090.

Version 2.5 and later, 3.0 included, can't be downloaded, so they can't run locally. The models you can run yourself are also the silent ones: built-in sound arrived with 2.5.

Wan vs Seedance vs Hailuo vs Veo

• Long single takes in Full HD: Wan 3.0 goes up to 30 seconds even at 1080p. Seedance 2.5 also reaches 30 seconds, but it's capped at 25 seconds in Full HD and doesn't accept photos with a real person's face.
• Cheap, fast short clips: Hailuo AI (MiniMax H3 Max Turbo), 5 to 15 seconds.
• The cheapest video to try: Veo 3.1 Lite on the Google Veo page, 12 credits for a 4-second silent 720p clip; Gemini Omni Flash is on the same page.
• Anime, stop-motion, or miniature styles: describe the style in your prompt and try it; results vary by model.

Sources

wan.video and its pricing page, the 2.1 and 2.2 GitHub repositories, Alibaba Cloud's release posts for 2.5, 2.6, 2.7, and 3.0, and the Wan-AI account on Hugging Face.

Klakix isn't affiliated with Alibaba; we provide access to this model through an API provider. Information is current as of Sep 24, 2026.