全部教程

Is Wan 3.0 Open Source? Can You Run It Locally?

No: Wan 3.0 has no public weights, so it does not run in ComfyUI or on your GPU. What is actually available, what it costs per second, and what we measured on our own runs.

2026年11月11日Wan3AIWan3AI
Is Wan 3.0 Open Source? Can You Run It Locally?

No. Alibaba has not released weights for Wan 3.0. It is offered only as a hosted model through Alibaba's own cloud and API resellers, so you cannot download it, load it in ComfyUI or run it on your own GPU. The last Wan flagship with public weights was Wan 2.2 (July 2025). If you need to run video generation locally, Wan 2.2 is still the Wan model to use; if you want Wan 3.0's 30-second clips, editing and extension, you have to call it over an API.

The confusion is understandable. Wan 2.1 and 2.2 were released with open weights, so many people expected the same for version 3. Searches like "wan 3.0 gguf", "wan 3.0 huggingface" and "wan 3.0 comfyui workflow" are common, and a few blogs and repositories answer them as if weights existed. They don't. Below is what is actually available as of November 11, 2026, how much it costs, and what we saw when we ran it on Wan3AI.

What Alibaba actually released

Detail
Public beta6 August 2026, on Alibaba Cloud Model Studio and Qwen Cloud (TNW)
General availability24 August 2026, announced by @Alibaba_Wan
AccessHosted API only, model ID wan3.0-video (Qwen Cloud model page)
TiersStandard and a faster, dearer Prime (The Decoder)
WeightsNone published

Alibaba describes Wan 3.0 as an all-in-one video model: text to video, image to video, first and last frame, references (images, clips and sound), editing and extension, with sound generated in the same pass and clips of up to 30 seconds. Those are Alibaba's claims; no independent benchmarks were published at launch.

Why "Wan 3.0 on Hugging Face" results mislead

Three things show up when you search for local Wan 3.0:

  1. Repositories that use the name but contain no model. One GitHub repository styled as an official Alibaba Cloud account holds only a README and a licence file; it belongs to a personal account, not to Alibaba.
  2. Blogs that claim Apache-2.0 weights. A few sites say Wan 3.0 weights are out under Apache 2.0. No Alibaba source says so, and there are no files to download. InstaSD's write-up walks through what is real in "Wan 3.0 ComfyUI" posts.
  3. ComfyUI nodes that call an API. Some "Wan 3.0" nodes are thin wrappers around a paid hosted endpoint. They run inside ComfyUI, but the video is generated on someone else's server and billed per second.

A quick test for any "local Wan 3.0" claim: look for a model file of several gigabytes. If the only thing you download is a node or a script asking for an API key, it is not local.

Your realistic options

1. Run Wan 2.2 locally. It is open, has mature ComfyUI workflows and runs on consumer GPUs with quantised builds. It is shorter (clips of a few seconds), has no native sound and no built-in editing or extension, but it costs nothing per clip once you own the hardware.

2. Call Wan 3.0 through an API. Alibaba, WaveSpeed, fal and Runware all list it at the same per-second price for Standard:

ResolutionPer second5-second clip30-second clip
480p$0.05$0.25$1.50
720p$0.10$0.50$3.00
1080p$0.20$1.00$6.00

Sources: WaveSpeed, fal, Runware. Prime costs about 1.5× on WaveSpeed. Endpoints differ by platform: WaveSpeed has separate video-edit and video-extend endpoints, fal's Wan 3 has neither at the time of writing.

3. Use a site that wraps the API. That is what Wan3AI is: an independent site (not affiliated with Alibaba) that sends requests to Wan 3.0 through WaveSpeed and charges credits at 10, 20 or 40 credits a second for 480p, 720p and 1080p.

What we measured

We ran each Wan 3.0 job type once on WaveSpeed on November 11, 2026 and used the results as the examples on our homepage. These are single runs, not a benchmark:

JobSettingsTimeCost
Text to video720p, 5 s, 16:9100 s$0.50 list
Image to video480p, 4 s64 s$0.20
First and last frame480p, 5 s67 s$0.25
Two reference images480p, 4 s136 s$0.20
Video edit (3 s in, 3 s out)480p174 s$0.30
Video extend (+4 s)480p84 s$0.20

Every clip came back with a soundtrack. The billing matched the published rules: an edit is billed for the seconds read plus the seconds made, an extension only for the new seconds.

The greenhouse shot from our text-to-video test: Wan 3.0, 720p, rendered in 100 seconds

This is the opening shot we generated from one paragraph. The prompt asked for a single continuous dolly-in; Wan 3.0 added a cut to a closer angle in the last second, a habit reviewers have also reported.

Text to video, 720p

A glass greenhouse on a sea cliff at dawn in late autumn. Mist rolls over the grass, orange leaves drift past. A gardener in a yellow raincoat, seen from behind, walks up and opens the glass door; warm light glows from the plants inside. Slow dolly-in, cinematic, natural colour. Sound: wind, distant surf, the door creaking open. No text, no captions, no logos, no brand names, no watermarks.

试试这个提示词

A paper crane taking off from a desk, a Wan 3.0 image-to-video test at 480p

For a cheaper test, the same model at 480p turns a still into a short clip. This one started from a photo of an origami crane.

Image to video idea, 480p

A small origami crane folded from pale blue paper sits on a sunlit wooden desk beside an open window. The crane comes to life: it unfolds its wings, flaps twice and glides out through the window into the morning sky. Gentle camera pan following it. Sound: paper rustling, birdsong outside. No text, no logos.

试试这个提示词

Limits worth knowing before you pay

  • Unrequested cuts. Wan 3.0 sometimes switches angle within a clip. One action per shot and an explicit "one continuous shot" help, but do not remove the risk.
  • Speech. Lip sync is loose in reviews we read and in Alibaba's own wording ("audio texture … still need improvement", via Zawya).
  • Wait times. Our runs took one to three minutes. Prime is faster at a higher price.
  • Rights. Commercial use follows Alibaba's and the reseller's terms. Keep real people's faces and other companies' logos out of prompts and uploads.

FAQ

Is Wan 3.0 free?

Not as open weights, and API access is paid per second. Some platforms give sign-up credits; on Wan3AI new accounts get 10 credits for the image tools, while the shortest Wan 3.0 clip (2 seconds at 480p) costs 20.

Will Wan 3.0 become open source?

Alibaba has not announced a release. Earlier Wan versions were opened months after their debut, and Wan 2.5 through 2.7 were never opened, so there is no reliable pattern to predict from.

Can I use Wan 3.0 in ComfyUI?

Only through API nodes that send the job to a hosted endpoint and bill per second. There is no checkpoint to load.

What are the hardware requirements for Wan 3.0?

None on your side: it runs on the provider's servers. For local generation, Wan 2.2's requirements apply.

Wan 3.0 or Wan 3.0 Prime?

Same parameters and features; Prime is the faster variant at about 1.5× the price. WaveSpeed lists about 77 seconds for a Prime text-to-video run against about 104 for Standard.