Wan 3.0 Turns a PDF Into 30 Seconds of Film. No Weights This Time.

TL;DR

On August 24 Alibaba took Wan 3.0 out of public beta, 18 days after the beta opened on August 6. The headline capability is 30 seconds of continuous video in a single pass, up to 1080p, with audio generated in that same pass, and inputs that now include PDFs, spreadsheets, slide decks and live web pages. It is an API. Alibaba has never published weights above Wan 2.2, and 3.0 does not break the streak. At $0.20 per second at 1080p, one 30-second render costs $6.00.

What actually shipped

Reuters reported the general rollout on Monday, quoting Alibaba Cloud's own WeChat announcement: the model generates 30-second videos from documents, spreadsheets, slides and web pages. Alibaba says that since the August 6 beta it has been used in short drama and film production, advertising and marketing, tourism promotion, and music video work.

The verified specification, per the model listing on fal and Alibaba Cloud's own pricing tiers:

• Duration: 2 to 30 seconds in one pass, and the model can pick the length itself from the prompt.
• Resolution: 480p, 720p or 1080p. There are exactly three tiers, which is the cleanest available proof that the widely repeated "native 4K" claim is not a thing. A vendor that shipped 4K would bill for 4K.
• Audio: generated alongside the video by default, not bolted on afterwards.
• Modes: text-to-video, image-to-video, and reference-to-video conditioned on images, video, audio, documents or webpages.
• Thinking mode: optional, lets the model reason about composition and motion before it renders. Useful for multi-event sequences and document references.
• Aspect ratios: 16:9, 4:3, 1:1, 3:4, 9:16, or model-selected.

The predecessor, Wan 2.7-Video, capped at 15 seconds. Doubling that is the entire structural claim here.

Audio in the same pass is the part worth understanding

Most video models are a silent film crew. They shoot the picture, and then a separate model scores it, dubs the dialogue and tries to make the lips land. Generating audio in the same pass is closer to shooting with a live boom mic on set: the mouth movement and the waveform come out of one process, so they are consistent by construction rather than by post-production luck.

This is also why the 30-second figure matters more than it looks. Anyone can chain three 10-second clips. What you cannot easily chain is coherent audio, continuous character identity and a single unbroken camera logic across the seam. Atlas Cloud's testing of Wan 2.7 found the chaining path actively hostile: source clips had to be 2 to 10 seconds, output capped at 15, and extending a clip made the model loop backwards instead of continuing forward.

What it costs

Pricing is per second of generated video, and it is linear, so there is no volume mercy at the long end. A 30-second 1080p render is $6.00. A 480p draft of the same length is $1.50.

Six dollars is a coffee, right up until you remember that generative video is a slot machine and the first four pulls are rarely the keeper. Budget for the reroll rate, not the render. The practical workflow is obvious and unglamorous: iterate prompts at 480p for $1.50 a go, then spend the $6.00 exactly once.

Access is application-gated on Alibaba Cloud Model Studio and Qwen Cloud under the model id wan3.0-video. Atlas Cloud reports the beta limits as 2 concurrent requests, 30 requests per minute and a 50-task async queue, which is a rate ceiling to plan around if you were imagining a batch pipeline.

The open-weight ceiling has not moved

This is the part that matters most if you run video locally. Wan built its reputation on genuinely open releases: Wan 2.2 shipped five model variants under Apache 2.0 on both Hugging Face and ModelScope, including a 5B text-and-image-to-video model small enough for a single consumer card. That is what made Wan the default backbone for a lot of ComfyUI work.

Everything after 2.2 has been API only. The Wan-AI organization on Hugging Face still has nothing numbered above 2.2, and 3.0 is a gated endpoint, not a download.

The nearest open alternative is MiniMax H3, a 33B dense omni-modal transformer whose weights went up on August 3. It does 4 to 15 second clips at up to 2K and 24fps with native stereo audio at 32 kHz, and dialogue in 11 stable languages. It is a serious release. It also carries a Community License with an application requirement for the United States, the EU, the UK and South Korea, so "open" comes with a postcode check.

The money underneath

The launch landed the same week Alibaba priced a HK$80 billion share placement, about $10.2 billion, selling 710 million ordinary shares at HK$112.70, a 3.6% discount to the prior close. Reuters called it the largest primary follow-on offering ever by a Hong Kong-listed company. Alibaba says 100% of net proceeds go to full-stack AI capabilities, including infrastructure.

The context for that raise is less cheerful: quarterly net profit fell 75% year on year on AI capital expenditure, the company has already spent nearly half of its three-year capex plan, and it now projects payback on AI investment at 2.5 years rather than 3. Wan 3.0 is one of the products that projection has to be right about.

What this means if you make things

For anyone producing short-form video, the document-to-video path is the genuinely new primitive. Not "describe a scene", but "here is the product spec sheet, here is the landing page, give me the 30-second explainer with a voice track". That collapses a storyboard, a script pass and an audio pass into one API call, and it is aimed squarely at ad and marketing volume work.

For anyone building on local inference, the read is colder. The best open video weights still top out around 15 seconds, and the vendor that popularized open video weights has now kept three consecutive releases behind a gate. Plan your pipeline for an API dependency, or plan it for 15 seconds.

Key Takeaways

• Wan 3.0 went generally available on August 24, 18 days after its August 6 public beta, with 2 to 30 second single-pass generation at up to 1080p.
• Audio is generated in the same pass as video, which is what makes the 30-second length usable rather than just long.
• Reference-to-video accepts PDFs, spreadsheets, slide decks and live webpages, aimed at explainer and marketing production.
• Pricing is linear per second: $0.05, $0.10 and $0.20 for 480p, 720p and 1080p. One 30-second 1080p render is $6.00.
• No weights. Wan 2.2 under Apache 2.0 remains the open ceiling, and there is nothing above it on the Wan-AI Hugging Face org.
• The nearest open substitute, MiniMax H3, does 15 seconds with native audio but restricts local deployment rights in the US, EU, UK and South Korea.

Sources: Reuters, "Alibaba launches Wan3.0 AI video model after $10 billion share sale", Alibaba Cloud, "Wan3.0: 30-Second AI Video Generation from Any Input", fal, Wan 3 model page, GitHub, Wan-Video/Wan2.2, Hugging Face, Wan-AI organization, Hugging Face, MiniMaxAI/MiniMax-H3, Atlas Cloud, "Wan 3.0 Preview: No 4K, No Weights, and the 15-Second Wall", TechNode Global on the Wan3.0 public beta, Reuters on the $10.2 billion Hong Kong share placement.