Wan 3.0 is released, and it has no open weights. Alibaba opened a public beta in early August 2026 and says it announced general availability on August 24. The model makes 2 to 30 second clips at up to 1080p with dialogue, music and sound effects. Unlike Wan 2.2, which is Apache 2.0, Wan 3.0 runs only on Alibaba's servers: in the Wan app, through the Model Studio API at $0.05 to $0.20 a second at list price, and through ComfyUI Partner Nodes. On 13 October 2026 neither Hugging Face nor GitHub held any Wan 3.0 files. Wan 2.2 users should keep it for local and LoRA work and call Wan 3.0 for shots that need its length or sound.
Wan 3.0 is out, but there are no open weights: Alibaba released it as a hosted model in August 2026, and on 13 October 2026 there was still no Wan 3.0 checkpoint, inference code or model license to download. You can use it in the Wan app, through Alibaba Cloud's Model Studio API, or through ComfyUI's paid Partner Nodes. Wan 2.2, released under Apache 2.0 in July 2025, remains the newest general-purpose Wan model you can run yourself.
Read on 13 October 2026: Alibaba Cloud's beta, launch and general availability posts, its Model Studio guide, API reference and pricing, the Wan app plans, the Wan-AI page on Hugging Face, the Wan-Video organization on GitHub and Comfy's Wan 3.0 guide. Prices are US dollars. We did not generate clips. To choose a video model for a persona, start with the AI influencer video guide.
Wan 3.0 release date
There are three dates, and all are in August 2026. Alibaba ran a short public beta, published prices a week later, and then declared general availability.
- August 6 to 7, 2026Public beta
Alibaba announces a beta; testers apply on Model Studio or Qwen Cloud. Clips up to 30 seconds, about double Wan 2.7.
- August 13, 2026Model Studio launch post with API prices
Per-second prices of $0.05, $0.10 and $0.20 for 480P, 720P and 1080P.
- August 24, 2026General availability
Alibaba Cloud says it announced general availability across its international channels, with a time-limited launch offer.
- October 11, 2026Docs last updated
The API reference still reads "Currently in preview." No weights have been published.
The preview label and the general availability post disagree, and both are Alibaba's own pages. In practice the model is on sale to anyone with an Alibaba Cloud account, so treat "preview" as a warning that parameters may still change.
What Wan 3.0 does
Wan 3.0 is one model for text to video, image to video and reference to video, with sound. From Alibaba's Model Studio guide:
- Length: 2 to 30 seconds per generation, or a smart duration setting that lets the model choose.
- Output: MP4 at 30 fps in 480P, 720P or 1080P.
- Audio: dialogue, background music and sound effects, on by default. You can turn audio off.
- Inputs: text; a first frame, or first and last frames; or up to 20 references per request, including up to 10 images, 5 video clips and 5 audio clips (video and audio each 15 seconds in total), plus one document or one web link.
- Editing and extension: both start from a reference video. For extension, input plus output must total 30 seconds or less.
- Two models:
wan3.0-videoandwan3.0-video-prime, which is priced about 40% higher.
Alibaba names its own weak spots in the launch post: "audio texture and on-screen text rendering accuracy are improving but not yet where we want them." One practical limit from the API reference: the video URL a task returns is valid for 24 hours, so download each result when it finishes.
Does Wan 3.0 have open weights?
No, and nothing official says it will. Open weights from the Wan team have always appeared in two places, and on 13 October 2026 neither had Wan 3.0.
- Hugging Face, Wan-AI: 27 models. None is Wan 2.5, 2.6, 2.7 or 3.0.
- Newest uploads there: the Wan2.2-Animate-2 checkpoints (July and August 2026) and Wan-Dancer-14B (July 2026).
- GitHub, Wan-Video: six repositories. Wan2.1, Wan2.2, Wan-Animate-2, Wan-Dancer, Wan-skills and a diffusers fork.
- Every repository in that organization is Apache 2.0. None is for Wan 3.0.
- Alibaba's beta, launch and general availability posts: access through Model Studio, no mention of weights or a license.
- Model Studio's wan3.0-video page marks model fine-tuning as unsupported.
The fine-tuning line is from Model Studio's wan3.0-video page. This is a pattern, not a delay. No Wan version after 2.2 is on the Hugging Face page, so Wan 2.2 is the last flagship Wan with downloadable weights. The team does keep releasing smaller open models: Wan Animate 2 arrived in August 2026 under Apache 2.0.
How to use Wan 3.0: app, API and ComfyUI
There are three official routes and they are billed separately. The Wan app's pricing page states that a subscription is for use on create.wan.video only and that the API needs its own purchase.
| Route | Price read on 13 October 2026 | Notes |
|---|---|---|
| Wan app (create.wan.video) | Free plan: credits from a daily check-in, one video at a time. Pro: $10 a month list, or $5 a month billed yearly, 300 credits. Premium: $40 list, or $20 billed yearly, 1,200 credits | Paid plans add 1080p, 10 to 30 second clips and watermark-free downloads. The subscription covers the app only |
| Model Studio API | wan3.0-video list price in Singapore: $0.05, $0.10 or $0.20 a second at 480P, 720P or 1080P, shown with a limited-time 30% discount. Prime: $0.068, $0.14 or $0.28 | Bills input video seconds plus output seconds. Failed requests are not billed. 30-second free quota in Singapore, valid 90 days |
| ComfyUI Partner Nodes | Comfy credits. The guide does not list a per-second price | Text, image and reference to video templates. Needs a Comfy login and the latest ComfyUI |
| Other APIs | Runway lists wan3 at 5, 10 or 20 credits a second for 480p, 720p or 1080p, at $0.01 a credit | The same list price as Alibaba, through an account you may already have |
Sources: the Wan app plans, Alibaba Cloud's Model Studio pricing, Comfy's Wan 3.0 Partner Node guide and Runway's API price list. Model Studio's US (Virginia) and Germany (Frankfurt) regions list lower prices, about $0.083 a second at 720P, with no free quota.
Source: Calculated: 10 seconds times the list price per second on Alibaba Cloud Model Studio pricing (Singapore, before the limited-time discount) and MiniMax API pricing, read 13 October 2026.
Alibaba's own example in its general availability post is a 30-second draft at 480P for about $1.50 and a 30-second finish at 1080P for about $6.00. Draft low, finish high.
Is Wan 3.0 free?
Only in small amounts. Two free allowances exist, and neither covers regular posting.
- Wan app free plan: credits from a daily check-in and one video at a time. The amount is not published. 1080p, 10 to 30 second clips and watermark-free downloads are listed on paid plans.
- Model Studio free quota: 30 seconds of combined input and output video per model in the Singapore region, valid for 90 days.
- No free local option: without weights you cannot run Wan 3.0 on your own GPU. Free local generation means Wan 2.2.
Free allowances on other tools are in our free AI video generator comparison.
What Wan 2.2 users should do
Keep Wan 2.2 and add Wan 3.0 as a paid tool for the shots that need it. The two do different jobs.
- Apache 2.0 weights, no per-second bill
- Your character LoRAs keep working
- 480p and 720p text-to-video and image-to-video models
- Wan Animate 2 for motion transfer, also Apache 2.0
- Hardware is the cost: the 5B model command needs 24GB of VRAM
- Single takes up to 30 seconds
- 1080P output at 30 fps
- Dialogue, music and sound effects in the same pass
- Up to 10 reference images instead of a LoRA
- Billed per second, and terms can change
A hybrid pipeline uses each where it is strong. Build the character once with a LoRA, keep everyday clips local, and send only the long or talking shot to the hosted model.
- 01Stills with a character LoRA
Local, in ComfyUI. A fictional adult persona you designed.
- 02Everyday motion on Wan 2.2
Short image-to-video clips with no per-clip cost.
- 03Long or talking shots on Wan 3.0
Pass your stills as reference images through the API or a Partner Node.
- 04Edit and label
Join the clips and add the AI label the platform asks for.
The pieces are covered in our LoRA training guide, the Wan 2.2 Animate tutorial, which covers both Animate versions in ComfyUI, and the best ComfyUI models guide for the other local video models. AI Influencers, one of our programs, has lessons on LoRA training for consistent faces and on image-to-video pipelines.
Wan 3.0 vs MiniMax H3
MiniMax H3 is the closest rival that does publish weights, with a license that limits where they can be used. Official specs side by side:
| Spec | Wan 3.0 | MiniMax H3 |
|---|---|---|
| Open weights | None | H3-Base checkpoints on Hugging Face, 768p output |
| License | No model license. Alibaba Cloud and Wan app terms apply | MiniMax H3 Community License Agreement, which excludes the US, EU, UK and South Korea |
| Clip length | 2 to 30 seconds | 4 to 15 seconds |
| Resolution | 480P, 720P or 1080P | 768p from the open weights. 2K needs the hosted stage |
| Audio | Dialogue, music and effects, on by default | 32 kHz stereo |
| API list price | $0.05, $0.10 or $0.20 a second at 480P, 720P or 1080P | $0.08 a second at 768p, $0.13 at 2K |
| ComfyUI | Paid Partner Nodes only | Local nodes for the open weights, where the license allows |
| Your own LoRAs | No fine-tuning: the model page marks it unsupported | Yes, locally |
Sources: the Alibaba pages above, MiniMax's H3 model card and API pricing.
Pick Wan 3.0 for a single take longer than 15 seconds, 1080p delivery or turning a document into video. Pick H3 for 2K output, or for local control outside the regions its license excludes; our MiniMax H3 ComfyUI guide covers that license in detail. Where each sits on the blind-vote leaderboards is in our AI video generator comparison.
How to prompt Wan 3.0
Alibaba's Wan 3.0 prompt guide gives one formula for a full prompt and says to skip any part you do not need:
Prompt = [Overall description]
+ [Reference material citation: Image N / Video N / Audio N]
+ [Shot N (start-end seconds): Subject + Scene + Motion + Aesthetic control]
+ [Dialogue: xx says: "content"]
+ [Sound effect / BGM]
+ [Style / Mood]
+ [Negative prompt list]- References are numbered by type in upload order: Image 1, Video 1, Audio 1. The guide warns that the numbering must match the upload order or the wrong material is cited.
- Shots connect end to end, with no gaps or overlaps, at 2 to 5 seconds each.
- Dialogue is written as the character, a speaking verb, a colon and the quoted line. If a shot should be silent, write "No dialogue", or the model may add some.
- Negative prompts list only what you do not want. The guide says not to pad the list.
- Prompt rewriting is on by default in the API, per the API reference. Turn it off when you need your exact wording.
Reference images and audio are how Wan 3.0 holds a character and a voice. Use them only for a persona you created or a person who gave written permission.
What changes often
The discount, first. Alibaba's pricing page marks the standard model "limited-time 30% off" without an end date, and the Wan app shows a 50% yearly discount. Weights are the other thing to watch: if Alibaba ever publishes them, they will appear on the Wan-AI page, and this page will say so.
Wan 3.0: FAQ
When was Wan 3.0 released?
Alibaba announced a public beta in the first week of August 2026, published a Model Studio launch post with API prices on August 13, and says it announced general availability across its international channels on August 24, 2026. Its API reference, last updated October 11, 2026, still labels the model "Currently in preview."
Is Wan 3.0 open source?
No. On 13 October 2026 Alibaba had published no Wan 3.0 weights, inference code or model license. The Wan-AI organization on Hugging Face listed 27 models, none of them Wan 2.5 or later, and the Wan-Video organization on GitHub had no Wan 3.0 repository. Wan 3.0 runs only on Alibaba's servers.
Is Wan 3.0 free?
Only in small amounts. The Wan app's free plan earns credits from a daily check-in, and Alibaba Cloud Model Studio gives a 30-second free quota in its Singapore region, valid for 90 days. After that you pay: the app's Pro plan lists at $10 a month, and the API at $0.05 to $0.20 a second at list price.
Can I run Wan 3.0 in ComfyUI?
Only through Partner Nodes, which need a Comfy account and credits and send the job to a hosted service. Comfy's guide lists text-to-video, image-to-video and reference-to-video templates. There is no Wan 3.0 model file to load locally. For local generation in ComfyUI, use Wan 2.2 or Wan Animate 2, both Apache 2.0.
Will Alibaba release Wan 3.0 weights?
Alibaba has not said. Its beta, launch and general availability posts describe access through Model Studio and say nothing about weights, a license or a repository. No Wan version after 2.2 appears on its Hugging Face page. The Wan team still publishes smaller Apache 2.0 models, most recently Wan Animate 2 in August 2026.
Wan 3.0 or MiniMax H3: which should I use?
Wan 3.0 for single clips longer than 15 seconds, 1080p output or a document as input. MiniMax H3 for 2K output through its hosted stage, or for local runs with open weights where its license allows. At list API prices, Wan 3.0 costs $0.10 a second at 720p and H3 $0.08 a second at 768p.