Wan 2.5 guide: features, access, cost and open-source status
Wan 2.5 is Alibaba's AI video model that makes 5 or 10 second clips with sound from text or an image. It is hosted-only: the newest Wan with downloadable weights is Wan 2.2 (Apache 2.0). Alibaba's API lists Wan 2.5 at $0.05 to $0.15 per second (checked 2026-10-27). Need local control, use Wan 2.2; otherwise use the API.
Key facts
| Wan 2.5 API price (Singapore, text-to-video) | $0.05/s at 480p, $0.10/s at 720p, $0.15/s at 1080p[1] |
|---|---|
| Wan 2.5 clip length and resolution | 5 or 10 seconds, up to 1920x1080[2] |
| Wan 2.5 audio | Generated by default; you can supply your own audio (WAV or MP3, 3-30 s)[2] |
| Newest open-weights Wan | Wan 2.2, Apache 2.0 licence[3] |
| Wan 2.5 launch | Wan2.5-Preview, reported 25 September 2025, not open source[6] |
Every fact here comes from Alibaba's own documents and the Wan repositories, with the date it was checked (2026-10-27).
What is Wan 2.5?
Wan 2.5 is Alibaba's video model that turns text or an image into a 5 or 10 second clip with synchronized sound. News coverage dated 25 September 2025 reported Alibaba launching it as Wan2.5-Preview.
Alibaba's API lists it as wan2.5-t2v-preview for text-to-video and wan2.5-i2v-preview for image-to-video. Both keep the "preview" label in the model name. Output goes up to 1920x1080, and prompts can run to 1,500 characters.
Is Wan 2.5 open source?
No. Wan 2.5 has no public weights, so you cannot download or self-host it.
Alibaba's Wan GitHub organization lists the Wan 2.1 and Wan 2.2 repositories. Its Hugging Face page shows no 2.5, 2.6, 2.7 or 3.0 weights. The Wan2.2 repository is under the Apache 2.0 licence. The launch report on Wan 2.5 states that, unlike earlier models, it is not open source.
A page that offers a local Wan 2.5 install is pointing at a hosted API node or a Wan 2.2 workflow. Check which model and provider the workflow calls.
Wan versions compared: open weights, audio and price
The newest Wan you can download is Wan 2.2. Every later version in Alibaba's docs is a hosted API model.
| Version | Weights | Audio | Alibaba API price per second (Singapore) |
|---|---|---|---|
| Wan 2.2 | Open, Apache 2.0 | Labelled silent in the hosted i2v models |
$0.02 (480p) to $0.10 (1080p) for wan2.2-t2v-plus |
| Wan 2.5 Preview | None published | Yes, by default | $0.05 / $0.10 / $0.15 at 480p / 720p / 1080p |
| Wan 2.6 | None published | Yes, by default | $0.10 (720p), $0.15 (1080p) |
| Wan 2.7 | None published | Yes (own audio supported) | $0.10 (720p), $0.15 (1080p) |
| Wan 3.0 | None published | Yes, native audio | $0.05 / $0.10 / $0.20 at 480p / 720p / 1080p, marked as a limited-time 30% off |
Notes on the table:
- "None published" means no weights on the Wan GitHub organization or the Wan-AI Hugging Face page.
- Wan 3.0 also has a
wan3.0-video-primemodel at $0.068, $0.14 and $0.28 per second. Wan 3.0 bills input and output video by duration. - The Wan 2.2 audio entry covers the hosted image-to-video models Alibaba labels as silent video. The open-weights Wan 2.2 family has a separate speech-to-video model. For character swaps with Wan 2.2, see the Wan 2.2 Animate guide.
How does Wan 2.5 differ from Wan 2.2 and Wan 3.0?
Wan 2.5 adds default audio generation to the hosted text-to-video line (see all Wan guides). Hosted Wan 2.2 already outputs 1080p, Wan 2.2 remains the open one, and Wan 3.0 is the newest hosted one.
Alibaba's documentation shows Wan 2.5 clips of only 5 or 10 seconds. Wan 2.6 and 2.7 accept any whole length from 2 to 15 seconds. Wan 2.7 also allows prompts up to 5,000 characters and removes the single or multi shot setting, so you describe shots in the prompt instead.
Alibaba's launch notice, dated 24 August 2026, says Wan 3.0 generates up to 30 seconds in one pass at up to 1080p, with native sound and picture. It accepts up to 10 images, 5 videos and 5 audio clips as references, and the API is open without an access application.
Wan 2.5 pricing: what a clip costs
At Alibaba's list prices, a 10-second Wan 2.5 clip costs $0.50 at 480p, $1.00 at 720p and $1.50 at 1080p. Prices are as of 2026-10-27.
| Resolution | Price per second | 5-second clip | 10-second clip |
|---|---|---|---|
| 480p | $0.05 | $0.25 | $0.50 |
| 720p | $0.10 | $0.50 | $1.00 |
| 1080p | $0.15 | $0.75 | $1.50 |
The sum is price per second times clip length. Alibaba states that only output is billed for these models, and its page does not list a separate audio charge. Alibaba's China (Beijing) region lists lower rates, for example $0.086 per second at 720p.
Retries change the real cost. If you keep one clip in three, a 10-second 720p clip costs about $3.00 in total.
Runway's API sells Wan 3.0 (wan3) at $0.05, $0.10 and $0.20 per second at 480p, 720p and 1080p, with credits at $0.01 each. It lists no Wan 2.5. For other models side by side, see the AI video cost calculator.
Wan 2.5 vs Kling: how do they compare?
Wan 2.5 is priced per second on a public API, and Kling is priced by plan or API, so the cheaper option depends on how you buy.
Use the numbers in the Kling AI pricing guide and compare them with the per-clip figures above. Compare at the same resolution, length and audio setting, and count retries on both.
Does Wan 2.5 generate audio?
Yes. Alibaba's API reference says the Wan 2.5 and Wan 2.6 series generate video with audio by default.
You can pass your own audio file (WAV or MP3, 3 to 30 seconds, up to 15 MB). Without one, the model generates matching background music or sound effects. Audio longer than the video is cut at the clip length, so a 5-second clip uses the first 5 seconds. Shorter audio leaves the rest of the clip silent: 3 seconds of audio in a 5-second clip gives 2 seconds of silence. A launch report noted drumming and music sometimes fell out of sync and that faces were not always consistent.
How can you use Wan 2.5?
You can call Wan 2.5 through Alibaba Cloud Model Studio or through third-party apps and APIs that host it.
Alibaba's free quota is 50 seconds per model in the Singapore region, valid for 90 days from activation. The API key, endpoint and model region must match. Third-party sites set their own prices and rules. If you want another way to make clips, see our Sora alternatives list.
Rules for faces and characters
Wan 2.5 animates supported images that pass the host's content checks; the image-to-video reference sets the format, size and file-size limits. Use faces and characters you have rights to. The Wan2.2 licence makes you responsible for how you use generated content, and Alibaba's Model Studio terms and any app you use add their own limits.
Questions
Is Wan 2.5 open source?
No. Alibaba has not published Wan 2.5 weights, and the official GitHub organization and Hugging Face model inventory contain no published Wan 2.5 weights as of 2026-10-27. Wan 2.2 is under Apache 2.0, which allows commercial use under that licence's terms.
Can I run Wan 2.5 locally or in ComfyUI?
Not with official weights, because none exist. A ComfyUI node can call a hosted API without running the model locally, so check which model and provider a workflow uses. For a local model, use Wan 2.2.
Does Wan 2.5 make audio?
Yes. Alibaba's API documentation says the Wan 2.5 and 2.6 series generate video with audio by default. Without an audio file it makes matching background music or sound effects.
Is there a free way to try Wan 2.5?
Alibaba's pricing page lists a free quota of 50 seconds per model for the Singapore region, valid for 90 days. That is five 10-second clips at any resolution. Third-party sites have their own free tiers and rules.
Is Wan 2.5 cheaper than Kling?
At $0.10 per second for 720p, a 10-second Wan 2.5 clip costs $1.00 on Alibaba's API. Kling pricing depends on the plan or API you use, so compare it using the Kling price guide linked in this article.
Sources
Related guides
- Kling AI pricing explained: what one usable clip costs in 2026Kling AI pricing turned into dollars per 5s and 10s clip, by resolution and audio, on every plan, plus what retries add. Checked 26 Oct 2026.
- AI video cost calculator: price per usable clip by modelWhat one usable AI video clip costs: Runway API rates for Kling, Veo, Wan and Seedance, plus retries, audio and reference-video billing. Checked 26 Oct 2026.
- Sora alternatives after the shutdown: 8 options, what they cost and what they replaceSora's app closed 26 April 2026 and its API on 24 September. Eight live options, their per-clip costs and which Sora strength each one covers.
More on Wan (open source)
- Wan 2.2 animation: how Animate and Replace modes work, what they need and what they costWan 2.2 Animate has two modes: animate a still, or replace a person in video. Input limits, failure causes and hosted prices from $0.04/s, checked 2026-10-26.
- ComfyUI Wan 2.2: how to run it locally, VRAM needs and GGUFRun Wan 2.2 in ComfyUI: official 5B needs 8 GB VRAM, 14B files add up to 28.6 GB in fp8. File sizes, GGUF quant table and setup steps, checked 2026-10-26.