Alibaba Wan3.0 Ships With 30-Second Clips and Closed Weights

Photo by Rezli on Unsplash

Alibaba released the Alibaba Wan3.0 AI video model on August 24, 2026, extending single-pass generation to 30 seconds at up to 1080p and 30fps, and shipping the weights closed after pre-announcing an open release. That single decision, more than the doubled clip length, is what the AI video community is arguing about this week.

Wan3.0 follows Wan2.7 as the next flagship from Alibaba's Tongyi Lab. It ships through Alibaba Cloud's Model Studio and Qwen Cloud, with time-based billing tied to output resolution rather than a per-clip flat fee.

The Open-Weight Reversal Nobody Should Have Been Surprised By

Alibaba pre-announced Wan3.0 as an Apache 2.0 release before launch. It shipped closed. This repeats a pattern across every Wan release since 2.1: Wan2.5 was pre-announced as open but weights never appeared, and Wan2.6 shipped as a closed commercial API with no open release. Wan2.7, released in March 2026, did eventually get select models under Apache 2.0, but only after six months of community pressure.

The strategic reversal is not subtle. Alibaba's first widely available video model, the open-weight Wan 2.1 family, launched in early 2025, and the shift to closed weights with Wan3.0 represents a notable pivot away from that posture. Community reaction, as tracked by wan27.org, has been described as caution driven by pattern recognition rather than outright anger. The people building on Wan already assumed the pre-announcement was aspirational.

That matters commercially because closed weights force every serious deployment through Alibaba Cloud's API, at Alibaba Cloud's prices, on Alibaba Cloud's uptime. The Wan3.0 closed weights open source debate is really a question about whether Alibaba's cloud economics now dominate its research posture. The Q1 fiscal 2026 numbers suggest they do.

Why Alibaba Cloud's Balance Sheet Explains the Model

The closed-weight decision reads differently once you look at what Alibaba is spending. Net income dropped 76% to 10.54 billion yuan in Alibaba's fiscal Q1 2026, while capital expenditure hit 67.68 billion yuan, a 75% year-over-year increase. Cloud and AI revenue is the offset the market is watching.

According to The Next Web, cloud and AI revenue grew 45% to 48.44 billion yuan in the June 2026 quarter, and AI model services alone now carry more than 16 billion yuan in annualised recurring revenue. Giving away a flagship video model as open weights would directly cannibalise the highest-growth line item Alibaba shows investors.

The financing side confirms the direction. Alibaba raised HK$80 billion, roughly $10.2 billion, in Hong Kong's largest follow-on share offering, and said it will direct all net proceeds into AI. The placement covered 710 million new ordinary shares priced at HK$112.70 each, a 3.6% discount to the previous close, sold to investors outside the US and expected to settle August 26, 2026. Alibaba stated the entire net proceeds will go into what it calls full stack AI capabilities spanning chips, infrastructure, and model development.

A company that just raised $10.2 billion for a full-stack AI push does not release the crown-jewel commercial model as Apache 2.0.

What Wan3.0 Actually Generates

The headline capability is the clip length. Wan3.0 generates clips up to 30 seconds in a single pass at resolutions up to 1080p, double the 15-second ceiling of predecessor Wan2.7. According to Dataconomy, the model runs at up to 30fps and Alibaba claims it maintains character detail, spatial layout, and motion consistency across the full clip length.

The input side is where Wan3.0 differs most from Western competitors. Alibaba Cloud announced via WeChat that Wan3.0 accepts documents, spreadsheets, slides, and web pages as inputs, not just text and images. Feeding a slide deck into a video model is not a use case Google, Runway, or OpenAI currently advertise. It is a direct pitch to enterprise marketing teams that already store product information in Office and Google Workspace files.

The model is also designed to produce synchronized micro-expressions and multilingual voice output. The multilingual audio is worth flagging because it puts Wan3.0 ahead of Veo 3.1's current shipping capabilities on that specific axis, according to Google's own DeepMind Veo page, which notes multilingual audio is still tested only against limited competing capabilities.

Alibaba lists confirmed real-world deployment across short drama and film production, marketing and advertising, tourism promotion, and music video creation. Independent benchmarks are absent. Dataconomy confirmed that no third-party performance benchmarks were published at launch and Alibaba has not released independent benchmark scores.

Wan3.0 vs Google Veo: The Clip-Length Gap Is Real

On duration alone, Wan3.0 vs Google Veo is not close. Veo 3.1 clips max out at 8 seconds per generation, with users able to request 4, 6, or 8 seconds. Extending a clip by adding 7 seconds per hop, up to 20 times, builds one continuous video up to roughly 148 seconds long, at 720p or 1080p and 24 fps. Wan3.0's 30-second single-pass output is nearly four times Veo 3.1's native ceiling.

That is not the whole comparison. Veo 3.1's differentiator is its unified audio-video approach: the model processes both modalities together using a joint diffusion process, dialogue matches lip movements with under 120ms accuracy per Google's own documentation, and ambient sounds respond to the visual environment. Wan3.0's synchronized micro-expressions and multilingual voice suggest a similar goal, but without published benchmarks it is a claim, not a measurement.

On resolution, the two are closer than the marketing suggests. According to MindStudio's Veo 3.1 breakdown, Veo 3.1 supports 720p, 1080p, and 4K, with 4K available through select platforms. Wan3.0 tops out at 1080p per Alibaba's own documentation. For enterprise short-form content on social platforms, that gap does not matter. For anything targeting broadcast, it does.

The pricing story is the most misleading part of any direct comparison. Alibaba Cloud's launch page advertises time-based billing with a temporary 30% discount on its standard tier through September 24, 2026. Reported pricing, via The Next Web, sits at $0.05 per second at 480p, $0.10 per second at 720p, and $0.20 per second at 1080p, with the caveat that failed generations are not billed. A 30-second 1080p clip therefore runs $6 before discount.

Cross-provider price comparisons require care because resolution, audio features, and output quality vary significantly between Wan3.0 and Google Veo 3.1. The not-billed-for-failed-generations clause is the interesting detail. Enterprise buyers running high-volume pipelines routinely eat 15 to 30 percent generation failure rates on current-generation video models, and that cost is normally passed through. Alibaba absorbing it changes the effective per-successful-clip cost more than the headline rate suggests.

The 18-Day Beta Nobody Talked About

A public beta version of Wan3.0 launched August 6, 2026, giving the model roughly 18 days of real-world use before the broader August 24 release. That is an unusually short public beta for a flagship generative video model, and it is not clear from Alibaba's communications whether the beta users were on the same 30-second, 30fps configuration that shipped at general availability.

An 18-day window suggests confidence, urgency, or both. It also means any early benchmarks or red-teaming from the community are working with a version that may have moved in the past three weeks. For teams evaluating Alibaba AI video generation 2026 options against Veo, Runway Gen-4, and Kling, that timing gap will matter when independent evaluations start appearing.

What the Infrastructure Spend Signals for the Next Twelve Months

Alibaba committed 380 billion yuan to AI infrastructure over three years in early 2025 and is reportedly weighing an increase to 480 billion yuan. The proposed 100 billion yuan uplift is the largest signal in the announcement, larger than any single Wan3.0 feature.

That spending pattern lines up with what other hyperscalers have done when moving from open research releases to closed commercial models. The DeepSeek V4 Pro pricing situation showed that even nominally open-weight Chinese models can face rapid commercial pressure. Wan3.0 is Alibaba deciding not to play that particular game with its flagship.

For competitors in the enterprise AI video segment, Wan3.0 sets a new duration floor at 30 seconds single-pass. Google and Runway will either extend native clip lengths or explain why 8-second extension chains produce better output than one continuous generation. Both companies have engineering timelines that suggest a response inside two quarters.

The Date to Watch

September 24, 2026, is when the 30% standard-tier discount ends. If Alibaba extends it, or bundles Wan3.0 access into existing Model Studio contracts, the pricing pressure on Veo and Runway sharpens. If the discount lapses and Wan3.0 pricing holds at the reported $0.20 per second at 1080p, expect Google to release a competitive per-second rate before Q4 earnings season. Watch the AI Industry news feed that week.