Alibaba Wan3.0 Launches 30-Second AI Video With Closed Weights

Alibaba's video bet

On August 24, Alibaba formally launched Wan3.0, the newest model in its Wan video-generation family. The model had been in an application-only public beta since August 6.37 The main feature is length. Wan3.0 can produce up to 30 seconds of video in a single pass, twice the 15-second ceiling of its predecessor, Wan2.7.26 That length matters, but the decision about its weights may matter more over time. Unlike Wan2.1 and Wan2.2, Wan3.0 has no publicly downloadable weights.1314

The timing was not accidental. The release came one day after Alibaba priced a share placement worth HK$80 billion, about $10.2 billion, and the company said all net proceeds would go toward full-stack AI.37 Reuters described the deal as the largest primary follow-on offering ever completed by a Hong Kong-listed company.5 The launch also arrived about a week after Alibaba reported a 75% drop in quarterly earnings, a decline driven by the same AI spending.6 Taken together, the share sale, the earnings hit and a paid flagship model tell one story: Alibaba is trying to turn heavy AI spending into revenue.

What Wan3.0 does

Alibaba's own materials describe the model as "everything to video." Along with text, images, audio and video, Wan3.0 accepts documents as input for the first time in the Wan line. Supported formats include PDF, Word, PowerPoint, Excel, Markdown, and Apple's Keynote, Pages and Numbers. Each request can include one file of up to 100MB and 50 pages.1 A single public webpage that doesn't require a login can also serve as source material.3 In reference mode, one request can combine up to ten images, five videos and five audio clips.310

Output runs at 30 frames per second in 480p, 720p or 1080p. When no input video is supplied, users can request any length from two to 30 seconds.3 Two features are meant to make the longer format usable. One suggests a clip length based on the prompt, and the other extends an existing generation.1 The editing tools introduced in Wan2.7 also carry over, so users can change visuals, plot and dialogue without regenerating the whole clip.13 Coverage also mentions synchronized facial micro-expressions and multilingual voice output.27

One small inconsistency stands out. Most reports say Wan2.7 is the model whose length Wan3.0 doubles, but The Decoder names Wan2.5 as the predecessor.106 Alibaba's own documentation names Wan2.7 as the baseline, which makes it the more reliable reference point.3

Pricing and positioning

Alibaba charges for each second of generated video. List prices are $0.05 per second at 480p, $0.10 at 720p and $0.20 at 1080p.1 At those rates, a full 30-second clip costs between $1.50 and $6.3 The Decoder also reports a faster "Prime" tier priced at $0.068, $0.14 and $0.28 per second, along with a temporary 30% discount on the Standard tier.10 TechNode reported that the discount ran on selected platforms from August 24 to September 23.8 Alibaba's release log lists wan3.0-video from August 6 and wan3.0-video-prime from August 20.20

Several outlets compared the pricing with Google's Veo 3.1. Its standard tier is quoted at $0.40 per second, double Wan3.0's top-resolution rate.27 This matches the pattern of Alibaba's Qwen language models, which have also undercut US competitors on price.2

Alibaba is aiming the model at business users. It names short dramas, film production, advertising, tourism promotion and music videos as uses since the beta opened.56 Technology.org argued that document input is the real differentiator, because it lets Wan3.0 work with material that companies already store on shared drives.6 That view seems right. Plenty of tools can make a 30-second clip from a prompt. Very few can turn a quarterly slide deck into a finished video in one API call.

Wan's open-weight era is over

This is the story's biggest implication. Alibaba built much of the Wan brand on open releases. Wan2.2 shipped under Apache 2.0 with weights and inference code, and it is still the license behind most free local Wan workflows.14 Dataconomy described closing Wan3.0's weights as a break from that open-source identity, timed to coincide with a push to make money from AI.7 Ground News's summary of the Reuters coverage described the same move from open-sourcing Wan2.2 to monetization.9

In practice, the shift began before Wan3.0. Several trackers report that Wan2.5, 2.6 and 2.7 were also API-only commercial models, and that Alibaba's published weights stop at Wan2.2.141819 A German-language review checked Hugging Face on October 1 and found that the newest Wan-AI uploads were still Wan2.2 variants, including Wan2.2-Animate-2-14B builds from July and August 2026.13 Winbuzzer reported that, as of August 25, Alibaba's official GitHub and Hugging Face organizations had no Wan3.0 model files.3 Developers can reach the model through Alibaba Cloud Model Studio, Qwen Cloud and wan.video, and third-party hosts such as fal and Replicate also carry it.1013

Not everyone agrees on this point. A GitHub repository under an "AlibabaCloud-Official" account calls Wan3.0 a model that "pushes the boundaries of open-source video synthesis."4 At least one tool directory states that the model ships with open weights.11 These claims don't hold up against repository checks by several independent outlets, all of which found no downloadable checkpoints.131417 One comparison guide warns that pages offering Wan3.0 downloads or ComfyUI workflows are not offering the real model.14 On the evidence available, Wan3.0 is a closed product. Any "open" label is either marketing or misinformation.

This leaves a gap in the open-weight video field. Current guides recommend Lightricks' LTX-2.5, MiniMax H3 and the aging Wan2.2 as the best options for running locally. Each comes with license strings attached: LTX-2.5 is free only for companies under $10 million in revenue, and MiniMax H3 excludes the US, EU, UK and South Korea.1319 Alibaba still releases weights for the previous generation, and one analyst suggests this pattern will likely continue: the flagship stays behind the API while older models stay open.13

Quality claims and caveats

The reporting disagrees on how well the model performs. At the formal launch, Winbuzzer said no neutral leaderboard had yet tested Wan3.0.3 By October, one roundup reported that it ranked first on Artificial Analysis's text-to-video leaderboards, both with audio (1,157) and without (1,128), and sixth for image-to-video with audio.19 If that holds, Wan3.0 is a frontier model on output quality, not just a cheaper option.

There are clear weak spots, though. Alibaba itself says audio texture and the accuracy of on-screen text still need work.3 Those are real limits for a model whose selling point is turning documents into videos with sound. Hosting provider Atlas Cloud ran 49 fixed-seed comparisons against Wan2.7 and ByteDance's Seedance 2.0. It found better identity retention and physics, but also repeated unwanted cuts in shots that were meant to be continuous.3 Atlas Cloud sells access to the models it tested, so it has a commercial interest.3 Still, unwanted cuts directly undercut the promise of a seamless 30-second take.

How it fits the wider release cycle

Wan3.0 lands in a crowded period for model releases. Anthropic has rolled out its Claude 5.5 family, with Opus 5.5 on September 22, Sonnet 5.5 on September 28 and Haiku 5.5 on October 7. All three offer 1M-token context windows.2425 On the Google side, Gemini 3.8 Flash arrived on September 2, and Gemini 4 Argon followed on September 30 in limited access.24 Nano Banana 2.1, Google's image model, launched on October 6.22 The two companies' product lines are also starting to overlap. Google's new Gemini work agent will offer Claude models alongside Gemini, initially for early-access customers.2326

Video is now a competitive front of its own. xAI released Grok Imagine Video 1.5 Lite on October 8.22 October roundups list Wan3.0 next to Seedance 2.5 and Gemini Omni Flash as the strongest hosted video models.19 All three are closed. Wan was once the main exception to that pattern, and it now follows the same pay-per-use model as its US rivals. Alibaba still competes hard on price, but its role as the leading open-weight video provider has largely ended.