# Wan (Alibaba)

> Alibaba's Wan video models: open-weight Wan 2.1 and 2.2 under Apache-2.0, the closed Wan 3.0 that tops the 2026 leaderboard, prices and rivals.

Source: https://metavert.io/wan  
Published: 2026-10-21  
Updated: 2026-10-21

**Wan** is Alibaba's family of video generation models, and it has two very different faces. Its open-weight releases, Wan 2.1 and Wan 2.2 (2025), are among the most widely used video models anyone can download and run, under the permissive Apache-2.0 license. Its newest model, Wan 3.0 (August 2026), is closed and sold through Alibaba Cloud, and as of October 20, 2026 it ranks first on Artificial Analysis's text-to-video-with-audio leaderboard. Wan comes from the same company as the [Qwen](https://metavert.io/alibaba-qwen) language models.

### The Open Wan Models

Alibaba released Wan 2.1 as open weights in February 2025 and Wan 2.2 on July 28, 2025, both under Apache-2.0, which lets anyone use, modify and redistribute them commercially with no revenue threshold or regional limits. Wan 2.2's flagship text-to-video and image-to-video models use a mixture-of-experts design, with one expert for early, high-noise denoising and another for detail: 27 billion parameters in total, 14 billion active per step, generating five-second clips at 480p or 720p. They need about 80 GB of GPU memory on a single card. A smaller dense model, TI2V-5B, generates 720p at 24 frames per second on a 24 GB consumer card such as an RTX 4090, which makes it one of the most accessible ways to try local video generation. The core models are silent.

Around them Alibaba has published specialized open models: first-and-last-frame interpolation, the VACE creation and editing model, speech-to-video (Wan2.2-S2V, August 2025) and character animation (Wan2.2-Animate, followed by Animate-2 in August 2026). The permissive license has made Wan a common base for community fine-tunes, LoRAs and ComfyUI workflows. Technically, Wan follows the [diffusion transformer](https://metavert.io/diffusion-transformers) design behind most current video models, paired with its own video autoencoder.

### Wan 3.0: Closed and at the Top

Since Wan 2.2, Alibaba's general-purpose releases (2.5, 2.6, 2.7 and 3.0) have been API-only, and its Hugging Face organization has no repositories for them. Some sites have claimed Wan 3.0's weights were released; they were not. Wan 3.0, released on August 13, 2026 according to Alibaba, generates up to 30 seconds in one pass, with extension, at 480p, 720p or 1080p, with audio including multilingual voice, although Alibaba itself has said the audio texture is "not yet where we want" it. It accepts an unusual range of inputs: text, images, audio, video and documents such as PDFs, slide decks, spreadsheets and Markdown.

Alibaba Cloud Model Studio publishes international list prices of $0.05, $0.10 and $0.20 a second at 480p, 720p and 1080p, with a Prime variant at $0.068 to $0.28 a second, and offers endpoints in Singapore, Frankfurt, Virginia, Beijing and Hong Kong (not every variant in every region). [Runway](https://metavert.io/runway) also hosts Wan 3.0, and [Higgsfield](https://metavert.io/higgsfield) lists Wan 2.6.

### How It Compares

Wan 3.0 led the text-to-video-with-audio board at an Elo of 1173 on October 20, 2026, narrowly ahead of Utopai X and ByteDance's [Seedance](https://metavert.io/seedance) 2.5, a spread small enough to call the top a near-tie. Against Seedance, Wan offers published prices and document inputs, while Seedance offers deeper reference control and more mature lip-sync. Among open models, Wan 2.2 competes with [Lightricks](https://metavert.io/lightricks)' LTX-2.5, which is newer and generates audio but carries a revenue-threshold license, and with Tencent's [HunyuanVideo](https://metavert.io/hunyuanvideo) 1.5, whose license excludes the EU, UK and South Korea. That makes Wan 2.2 the cleanest license in open video generation as of October 2026. See [Wan vs Seedance](https://metavert.io/compare/wan-vs-seedance), [LTX vs Wan](https://metavert.io/compare/ltx-vs-wan), [Wan vs HunyuanVideo](https://metavert.io/compare/wan-vs-hunyuanvideo), [open-weight models](https://metavert.io/open-weight-models) and [text-to-video](https://metavert.io/text-to-video).

## Related Topics

- [Wan vs Seedance](https://metavert.io/compare/wan-vs-seedance) — The top of the 2026 leaderboard
- [LTX vs Wan](https://metavert.io/compare/ltx-vs-wan) — Open-weight video models for commercial use
- [Wan vs HunyuanVideo](https://metavert.io/compare/wan-vs-hunyuanvideo) — Chinese open-weight lineages and license geography
- [Alibaba Qwen](https://metavert.io/alibaba-qwen) — Alibaba's language model family
- [Text-to-Video](https://metavert.io/text-to-video) — How video generation models work
- [Open-Weight Models](https://metavert.io/open-weight-models) — Models you can run yourself
- [Diffusion Transformers (DiT)](https://metavert.io/diffusion-transformers) — The architecture behind Wan
- [Seedance](https://metavert.io/seedance) — ByteDance's rival model family

## Further Reading

- [Wan 3.0: 30-second generation from any input](https://www.alibabacloud.com/blog/wan3-0-30-second-ai-video-generation-from-any-input_603452) — Alibaba Cloud, August 2026
- [Alibaba Cloud Model Studio pricing](https://www.alibabacloud.com/help/en/model-studio/model-pricing) — Wan 3.0 per-second rates and regions
- [Wan 2.2 on GitHub](https://github.com/Wan-Video/Wan2.2) — Models, hardware requirements and Apache-2.0 license
- [Wan-AI on Hugging Face](https://huggingface.co/Wan-AI) — The open Wan models
- [Wan: Open and Advanced Large-Scale Video Generative Models](https://arxiv.org/abs/2503.20314) — The Wan technical report, 2025
- [Alibaba releases Wan 3.0](https://thenextweb.com/news/alibaba-wan3-video-model-after-share-sale) — The Next Web; closed weights