Wan
Wan is Alibaba's family of AI video models; Wan 2.1 and Wan 2.2 are free open models under the Apache 2.0 license that run on a home graphics card, while the newer Wan 2.5 to 2.7 and Wan 3.0 are offered only through Alibaba's apps and API as of Nov 6, 2026.
Wan is Alibaba's AI video model family, and its open versions, Wan 2.1 (released Feb 25, 2025) and Wan 2.2 (Jul 28, 2025), are free to download and use commercially under the Apache 2.0 license: the small Wan 2.1 1.3B model needs only 8.19 GB of video memory, and Wan 2.2's 5B model runs on one 24 GB card such as an RTX 4090. The newer models are not open as of Nov 6, 2026: Wan 2.5, 2.6 and 2.7 are sold through Alibaba Cloud's API, and Wan 3.0, which ranks 8th on Arena's text-to-video leaderboard (Arena, read Nov 6, 2026), has no published weights on Alibaba's Hugging Face page.
Pick Wan 2.2 if you want video generation on your own computer, with no per-clip cost, no account and no content sent to anyone; it is also built into ComfyUI and Diffusers. Its quality is well behind the leaders (Wan 2.2 A14B is 44th on Arena), so for the best results use a hosted model such as Google's Gemini Omni in Google Flow or ByteDance's Seedance, and for a stronger open model with sound, MiniMax's H3 (11th), which needs a multi-GPU server. Run AI Locally covers the setup and Best AI Video Generator the hosted rivals.
Quick facts
| Get it | Weights on GitHub and Hugging Face; app at wan.video; Alibaba Cloud Model Studio API |
| Maker | Alibaba |
| Open models | Wan 2.1 (Feb 25, 2025), Wan 2.2 (Jul 28, 2025), Wan2.2-Animate-2 (Aug 2026) |
| API-only models | Wan 2.5 preview, 2.6, 2.7; Wan 3.0 |
| License | Apache 2.0 for the open models |
| Hardware | Wan 2.1 1.3B: 8.19 GB VRAM; Wan 2.2 5B: 24 GB; Wan 2.2 14B: 80 GB on one GPU |
| Ranking | Wan 3.0 8th (1477), Wan 2.7 22nd, Wan 2.2 A14B 44th on Arena text-to-video (read Nov 6, 2026) |
| Open weights | Yes for 2.1 and 2.2; no for 2.5 and later |
Models
Wan 2.1 (Feb 25, 2025) was Alibaba's first open release: text to video and image to video, with a 1.3B model small enough for 8.19 GB of video memory (GitHub).
Wan 2.2 (Jul 28, 2025) uses a mixture-of-experts design: the 14B models make each video in two passes, a high-noise expert for the rough structure and a low-noise expert for the detail. Its models are T2V-A14B and I2V-A14B (text or image to video at 480p and 720p), TI2V-5B (a dense 5B model doing both at 720p and 24 fps), S2V-14B for speech-driven video (Aug 26, 2025) and Animate-14B for animating or replacing a character from a reference video (Sep 19, 2025) (GitHub). Wan2.2-Animate-2 and Wan-Dancer followed in Jul and Aug 2026 (Hugging Face).
Wan 2.5, 2.6 and 2.7 are hosted only. In Alibaba Cloud's API, Wan 2.7 (snapshot dated Jun 12, 2026) and Wan 2.6 make clips of 2 to 15 seconds at 720p or 1080p with generated sound or your own audio track, and Wan 2.5 preview makes 5 or 10 seconds (Alibaba Cloud). Billing is per second, higher for higher resolution.
Wan 3.0 appears on Arena's leaderboard at 8th, and Pika offers it in its app, but Alibaba's API reference listed no Wan 3.0 model and its Hugging Face page showed no Wan 3.0 weights on Nov 6, 2026.
How to run Wan locally
Wan 2.2's code and weights are on GitHub and Hugging Face, and Alibaba integrated it into ComfyUI and Hugging Face Diffusers on release day (GitHub). In MDMZ's Aug 7, 2025 tutorial, ComfyUI's built-in templates offered two Wan 2.2 14B workflows, which he advised only with at least 24 GB of video memory, and a 5B workflow he said runs on as little as 8 GB, lower than the 24 GB the Wan 2.2 README gives for the 5B model. The practical choices by graphics card:
| Video memory | Model | What you get |
|---|---|---|
| 8 GB or more | Wan 2.1 T2V-1.3B | Text to video with the smallest model |
| 24 GB (RTX 4090 class) | Wan 2.2 TI2V-5B | 720p text or image to video at 24 fps |
| 80 GB per GPU, or several GPUs | Wan 2.2 A14B, S2V, Animate | The best open Wan quality, speech and character animation |
Source: Wan 2.1 and 2.2 READMEs, read Nov 6, 2026. Run AI Locally and Hugging Face explain the tools.
Wan vs the alternatives
MiniMax H3 is the strongest open-weight video model on Arena (11th, 1456, against Wan 2.2's 1132) and makes 15 seconds with stereo sound, but its 33B model and serving example need four GPUs and its license is MiniMax's own, not Apache 2.0.
Google Flow gives Gemini Omni 1.2 Flash, Arena's 1st, with 50 free credits a day, if you do not need to run anything yourself.
Seedance 2.0 and 2.5 (6th and 7th) are the closest hosted rivals to Wan 3.0.
Kling 4.0 is the hosted pick for 30-second single takes.
Frequently asked questions
Is Wan open source? Wan 2.1 and Wan 2.2 are open under the Apache 2.0 license, which allows commercial use; Wan 2.5, 2.6, 2.7 and Wan 3.0 have no published weights as of Nov 6, 2026.
Can Wan run on my PC? Yes. Wan 2.1's 1.3B model needs about 8 GB of video memory, and Wan 2.2's 5B model runs on one 24 GB card such as an RTX 4090; the 14B models need 80 GB on a single GPU.
What is the newest Wan model? Wan 3.0 is the newest Wan model on Arena's leaderboard, at 8th as of Nov 6, 2026; the newest in Alibaba Cloud's API is Wan 2.7, and the newest open one is Wan 2.2.
How long are Wan videos? Wan 2.6 and 2.7 in Alibaba Cloud's API make 2 to 15 seconds; the open Wan 2.2 models make short 480p or 720p clips.
Is Wan free? The open Wan 2.1 and 2.2 models are free to download and run on your own hardware; the API models are billed per second of video.
Sources
- Wan2.1 and Wan2.2, Wan-Video on GitHub, read Nov 6, 2026
- Wan-Video organization, GitHub, read Nov 6, 2026
- Wan-AI models, Hugging Face, read Nov 6, 2026
- Text-to-video API reference, Alibaba Cloud Model Studio, read Nov 6, 2026
- Wan, Alibaba, read Nov 6, 2026
- Wan 2.2, the generative AI model built for the language of film, Alibaba Cloud, Jul 28, 2025 (video)
- How To Use WAN 2.2 in ComfyUI: The BEST FREE AI Video Model, MDMZ, Aug 7, 2025 (video)
- Text-to-video leaderboard, Arena, read Nov 6, 2026
Moments
Wan 2.2, the generative AI model built for the language of film
Alibaba Cloud
How To Use WAN 2.2 in ComfyUI: The BEST FREE AI Video Model
MDMZ
- Pick Wan 2.2 if you want video generation on your own computer, with no per-clip cost, no account and no content sent…
- Wan 2.2 (Jul 28, 2025) uses a mixture-of-experts design: the 14B models make each video in two passes, a high-noise…
- Wan 2.2's code and weights are on GitHub and Hugging Face, and Alibaba integrated it into ComfyUI and Hugging Face…
Photos
See also
-
Google FlowGoogle Flow is Google's AI filmmaking app at flow.google.com, built on the Gemini Omni Flash and Veo video models and Nano Banana for images. Everyone 18 or older in a supported country gets 50 free credits a day; Google AI Pro adds 1,000 credits a month and Ultra 10,000 or 25,000, as of Nov 6, 2026. -
SeedanceSeedance is ByteDance's family of AI video models; Seedance 2.0 and 2.5 rank 6th and 7th on Arena's text-to-video leaderboard as of Nov 6, 2026, and 2.5 makes up to 30 seconds with sound in one pass. You can use it in ByteDance's Dola and Dreamina, through the BytePlus API, or inside Runway, Luma, Pika and HeyGen.
-
HailuoHailuo is MiniMax's AI video app at hailuoai.video; its current model, MiniMax H3 (Jul 31, 2026), makes 4 to 15 second videos with stereo sound at up to 2K and has open weights. H3 ranks 11th on Arena's text-to-video leaderboard as of Nov 6, 2026.
-
Run AI LocallyTo run AI on your own computer, install Ollama or LM Studio (both free) and download an open model sized to your memory, such as Gemma 4, Qwen 3.5 or 3.8, or gpt-oss 20b. With 8 to 16 GB a small model runs well; a 24 GB graphics card or a 32 GB Mac runs 27B to 31B models.
-
Best AI Video GeneratorThe best AI video generator in Nov 2026 on LMArena's votes is Google's Gemini Omni Flash, and on Artificial Analysis's it is Alibaba's Wan 3.0; ByteDance's Seedance and MiniMax's Hailuo models rank near the top of both. OpenAI closed the Sora app in 2026, and Sora 2 and Veo 3.1 now rank 13th and 14th on LMArena.
-
PikaPika is an AI video app from Pika Labs whose newest model, Pika H3 Flex (Oct 29, 2026), makes up to 15 seconds of video with stereo sound and has open weights. Pika has a free plan that runs on bought credits, and paid plans cost $10, $35 or from $95 a month as of Nov 6, 2026.