What is Wan?
Wan is Alibaba’s family of AI video and image generation models, offered to the public through the wan.video creative platform. The site lists text-to-video, image-to-video, text-to-image, image-to-image and image editing, and Alibaba’s December 2025 announcement of Wan2.6 names wan.video as the official Wan website alongside Model Studio and the Qwen App. Recent hosted versions generate clips with matching speech, music and sound effects.
The series has two tracks. Wan2.2, released on GitHub in July 2025 under the Apache 2.0 licence, includes open text-to-video, image-to-video, speech-to-video and character animation models that you can run on your own GPUs. Later versions, including Wan2.5, Wan2.6 and the Wan 2.7 suite that Runpod reports Alibaba released in April 2026, are available only through hosted services. In October 2026 the platform was live, and EvoLink reported that a Wan 3.0 video model had entered invite testing on Alibaba Cloud in August 2026.
Key Features
- Text and image to video: Wan generates video clips from a written prompt or from a still image you upload, according to the wan.video site. The same platform also creates and edits images, so storyboards and final clips can come from one account.
- Native audio and lip sync: Wan produces synchronised voices, music and sound effects with the video in its hosted versions. The Decoder reported that Wan2.5-Preview generated 10-second 1080p clips with matching audio, and Alibaba says Wan2.6 improved audio-visual sync further.
- Reference-to-video casting: Wan2.6-R2V takes a reference clip of a person, animal or object and keeps its look and voice in new scenes, Alibaba says. The company pitches it as a way for users to star in their own AI-generated videos.
- Multi-shot storytelling: Wan2.6 builds short narratives across several shots while keeping subjects visually consistent, per Alibaba’s announcement. Clips run up to 15 seconds, which helps for ads, trailers and social stories that need more than a single camera angle.
- Character animation and replacement: Wan2.2-Animate animates a character image with the motion from a source video, or swaps that character into the video in place of the original person. The model weights are public on GitHub and Hugging Face for self-hosting.
- Instruction-based video editing: Wan 2.7 edits existing clips through natural-language instructions covering actions, dialogue, scenes, style and camera work, Runpod reports. It can re-sync lip movement after dialogue changes and keep up to five characters consistent.
- Open-weight models: Wan publishes its Wan2.2 models under Apache 2.0, including T2V-A14B, I2V-A14B, TI2V-5B, S2V-14B and Animate-14B. The TI2V-5B model renders 720p at 24 fps on a single 24GB GPU such as an RTX 4090.
- Speech-to-video: Wan2.2-S2V-14B turns an image and an audio track into a talking or singing video, according to the GitHub release notes. A September 2025 update added CosyVoice text-to-speech support, so a script can drive the whole clip.
- API access through Alibaba Cloud: Wan models are offered to developers through Alibaba Cloud Model Studio, and third-party hosts resell access. The Decoder reported API prices of roughly 5 to 15 cents per second of video at the time of Wan2.5.
Pros
- Free entry point: Wan offers a free plan on wan.video, and Comparedge reports it allows generation without credits at one concurrent job. That lets you test prompts and styles before paying, which few hosted video models permit.
- Low reported subscription prices: Wan’s Pro plan is reported at $10 a month, or $5 a month billed yearly, for 300 credits. Trackers say that covers up to 60 videos, which is cheap compared with many closed video generators.
- Self-hosting option: Wan is one of the few major video model families with open weights under Apache 2.0, through Wan2.2. Studios with their own GPUs can run it privately, fine-tune it and use it in ComfyUI or Diffusers pipelines.
- Sound arrives with the picture: Wan’s hosted models create dialogue, music and effects alongside the video. Creators skip a separate audio pass for short social clips, although timing should still be checked before publishing.
- Wide set of inputs: Wan accepts text, images, audio and reference video across its model range. One family covers prompt-to-video, photo animation, talking avatars and character swaps, so teams can standardise on a single provider.
- Fast release cadence: Wan has moved from Wan2.2 in July 2025 to Wan2.6 in December 2025 and Wan 2.7 in April 2026, per GitHub, Alibaba and Runpod. Each release added features such as audio, multi-shot output or editing.
Cons
- Likeness tools raise consent issues: Wan can animate a photo or replace a person in a video, and Wan2.6 copies a reference subject’s look and voice. Alibaba’s announcement mentions no consent checks, so only use faces and voices you have permission to use.
- Newest models are not open: Wan publishes weights only up to Wan2.2, and Runpod found none for Wan2.5, 2.6 or 2.7. The features that matter most, such as audio and reference casting, are therefore available only on hosted services.
- Official prices are hard to read: Wan’s pricing page loads its figures in the browser, so we relied on Comparedge and other trackers. Earlier coverage cited plans from $6.50, which shows prices have changed, so check the live page before buying.
- Heavy hardware for local use: Wan’s 14-billion-parameter open models need at least 80GB of VRAM on a single GPU, according to the GitHub README. Only the smaller TI2V-5B fits a 24GB consumer card, which limits home self-hosting.
- Quality is still uneven: Wan outputs can drift, with The Decoder noting music and drumming falling out of sync and faces changing between shots in Wan2.5. CostBench also lists weaker character consistency in earlier versions.
- Strict refund terms: Wan credits are non-refundable, non-transferable and expire two years after redemption, Comparedge reports, and the first paid period is non-refundable. Test on the free plan first so you do not lock money into credits you will not use.
Who is Using Wan?
- Social media creators: Making short clips with built-in sound and dialogue.
- Marketers and agencies: Drafting multi-shot ads and product videos.
- AI developers: Self-hosting Wan2.2 or calling Wan through Alibaba Cloud.
- Animators and VFX artists: Animating characters and swapping performers in footage.
Pricing
Wan has a free plan on wan.video, with Pro and Premium subscriptions reported by Comparedge as of July 2026 (checked October 2026).
| Plan | Price | What you get |
|---|---|---|
| Free | $0 | Generation without credits, one video and one image job at a time, six image styles |
| Pro | $10/month, or $5/month billed yearly (reported) | 300 credits a month, watermark-free downloads, 1080p, 10 and 15 second videos, upscaling |
| Premium | $40/month, or $20/month billed yearly (reported) | 1,200 credits a month, eight concurrent video jobs, Pro features |
| Credit packs | From $1.50 for 30 credits to $100 for 3,900 (reported) | One-off credits valid for two years |
| Open Wan2.2 weights | Free, Apache 2.0 | Self-hosted models; you pay for your own GPUs |
Vijay Talks AI lists the same Pro and Premium yearly rates, while CostBench gives conflicting annual figures and The Decoder cited plans from $6.50 in 2025. The official page renders prices only in the browser, so open create.wan.video/pricing while signed in and confirm the current rate and credit cost per model before you subscribe.
What Makes Wan Unique?
Wan is unusual among leading video generators because the same family offers both free open weights and a polished hosted app with native audio. Developers can run Wan2.2 locally under Apache 2.0, while creators who want speech, multi-shot scenes or reference casting use the newer hosted versions on wan.video. Few rivals give you that choice between owning the model and renting the latest one from a single maker.
Compatibilities and Integrations
- Web platform: wan.video and create.wan.video in any modern browser.
- Developer access: Alibaba Cloud Model Studio API, plus third-party hosts that resell Wan models.
- Open models: Wan2.2 weights on GitHub and Hugging Face, with ComfyUI and Diffusers integration.
- Other apps: Alibaba says Wan2.6 is being integrated into the Qwen App.
Wan Tutorials
A first clip on the hosted platform takes only a few minutes.
- Go to wan.video and sign in to the creation workspace.
- Choose text-to-video, image-to-video or a reference-based mode.
- Write a prompt describing the subject, action, camera and sound you want.
- Pick resolution and length within your plan’s limits, then generate.
- Review the clip for sync and consistency, then download or regenerate.
How We Rated It
| Criterion | Score | Why |
|---|---|---|
| Accuracy and Reliability | 3.7 | Good clips, uneven sync |
| Ease of Use | 4.0 | Simple web workspace |
| Functionality and Features | 4.6 | Audio, R2V, editing, animation |
| Performance and Speed | 3.6 | Speed not independently tested |
| Customization and Flexibility | 4.4 | Open weights plus hosted modes |
| Data Privacy and Security | 3.2 | No published consent safeguards |
| Support and Resources | 3.6 | GitHub docs and model cards |
| Cost-Efficiency | 4.3 | Free plan, low reported prices |
| Integration Capabilities | 4.2 | API, ComfyUI, Diffusers |
| Overall Score | 4.0 | Average of the nine criteria |
Summary
Wan suits creators who want low-cost AI video with sound, and developers who value open weights they can run and adapt themselves. Its main limitations are that the newest audio and casting features are hosted only, and that likeness tools need careful consent handling. We rate it highly for features and value. Start on the free plan, test a short prompt with dialogue, and only upgrade once the clips meet your quality bar.
Reviews
Write a Review