Is Wan 3.0 open source?
No — and here is what shipped instead

That matters more than a label, because the three fail in different ways. No licence means self-hosting is not a technical problem you can solve with a bigger card — there is nothing to license. And the practical consequence is narrower than it sounds: you lose the ability to run it offline, not the right to sell the output. Commercial use travels with the API, which is why the one thing you genuinely cannot do is the one thing most people came here for. The full Wan 3.0 specification, row by row.

Published Wan 3.0 checkpoints

0

Published Wan 3.0 checkpoints

Model IDs, both API-only

2

Model IDs, both API-only

The last release with weights

Wan 2.2

The last release with weights

That release's licence, still valid

Apache 2.0

That release's licence, still valid

What the closed one costs to run

$0.05/s

What the closed one costs to run

• No Hugging Face checkpoint
• No GitHub weights repository
• No self-hosting licence
• No quantised or distilled build
• No announced plan to publish either

Checked by Hao Lin against the pages linked above on 14 September 2026. These are negative claims — "there is no X" ages faster than "there is an X" — so this table is re-checked weekly and the date moves when it does. Found a row that has gone stale? Tell us.

• Weights

  Wan 2.2 (open)

  Published, Apache 2.0

  Wan 3.0 (closed)

  None

• Released

  Wan 2.2 (open)

  28 Jul 2025

  Wan 3.0 (closed)

  GA 24 Aug 2026

• Runs offline

  Wan 2.2 (open)

  Yes, on your own card

  Wan 3.0 (closed)

  No

• Entry VRAM

  Wan 2.2 (open)

  ≈8 GB, TI2V-5B quantised

  Wan 3.0 (closed)

  None — nothing runs locally

• Native audio

  Wan 2.2 (open)

  No

  Wan 3.0 (closed)

  Yes, same pass

• Max single pass

  Wan 2.2 (open)

  5s typical

  Wan 3.0 (closed)

  2–30s

• Lowest tier

  Wan 2.2 (open)

  720P

  Wan 3.0 (closed)

  480P

• Document / webpage input

  Wan 2.2 (open)

  No

  Wan 3.0 (closed)

  Yes

• Ongoing cost

  Wan 2.2 (open)

  Your electricity and your card

  Wan 3.0 (closed)

  $0.05–$0.20 per generated second

• Rights over output

  Wan 2.2 (open)

  No revenue cap, no field limits

  Wan 3.0 (closed)

  Commercial use on paid plans

Wan 2.2 model IDs, for the search you are about to run: Wan2.2-T2V-A14B, Wan2.2-I2V-A14B, Wan2.2-TI2V-5B, Wan2.2-S2V-14B, Wan2.2-Animate-14B. The 5B build is the one that fits a consumer card.

• The page advertises 4K

  Wan 3.0 tops out at 1080P. There is no 4K tier and no way to request one through the API

  Anything labelled 4K is an upscaler running after generation — a bigger file, not more detail. Ask what the upstream resolution was

  Check the spec page first

• No model ID appears anywhere on a finished clip

  The API returns the model ID in the response. A site that does not print it has chosen not to, and the reason is rarely flattering

  Ask to see the raw response for one clip. It is one field, and there is no cost to showing it

  Printed on every clip here

• Maximum length is 15 seconds, or five

  Wan 3.0 does 2–30 seconds in one pass. Fifteen is the ceiling on reference video input, and five is roughly where the previous generation sat

  Ask for thirty seconds in one generation. An older engine cannot produce it, whatever the label says

  2–30s, any whole number

• The clip comes back silent

  Wan 3.0 writes the soundtrack in the same pass as the picture, and switching it off does not change the price. Silence usually means an older model with audio bolted on afterwards

  Generate a shot with a specific sound in the prompt and listen to whether it lands on the action

  Sound in the same pass

• "Free and unlimited, no watermark, forever"

  Alibaba meters this model per second of output from the first second. Nobody is absorbing an unlimited bill, so something else is happening — an older model, a queue that never clears, or a limit that appears after signup

  Find the per-clip or per-second number before you invest a prompt in it. A provider that will not say is telling you something

  One free clip, then metered

These checks are worth running on this site too, which is the point of printing the receipt above. Our answer to all five is on the page rather than in a support reply — and if one of them ever stops being true here, it will be because we broke something, not because we quietly changed engines.

There is no download, so the only remaining question is whose infrastructure you use. All three of these are legitimate — they differ in what they ask of you before the first clip, not in which model they reach.

1. 01

   Alibaba Cloud Model Studio — the source

   The first-party route, billed at $0.05 / $0.10 / $0.20 per generated second. It wants a cloud account, a workspace and a region that matches your key, and the new-account free quota is one-off and Beijing-only. Right answer if you are building a product on it.

2. 02

   An inference platform — if you are writing code anyway

   fal, Atlas Cloud, OpenRouter and others resell the same model behind their own SDKs. Worth it when you already run several models through one billing relationship and want one more. You are paying a margin for that convenience.

3. 03

   Here — if you want to see it work before any of that

   No cloud account, no key, no region to get wrong. Your first clip is free — 480P, up to three seconds, with sound, on the real wan3.0-video. An email, no card. When you want thirty seconds at 1080P, plans start at $12.90 a month billed yearly, and a failed run never costs a credit.

Take the free clip

All three reach the same weights, so the only real question is how much has to happen before your first clip does. Routes one and two ask for a cloud account, a key and a region that matches it — worth doing once you know the model earns its place in your work. This is the route that lets you find that out first, with your own shot rather than someone's showreel.