Alibaba Wan3.0 Video Model Doubles AI Clips to 30 Seconds

What Alibaba shipped

Alibaba has made Wan3.0 generally available. It is the newest model in its Wan video generation family, and one generation can now produce up to 30 seconds of video from text, images, audio, existing footage, or office documents. Alibaba Cloud announced the release on Monday, August 24, 2026, after a public beta that began on August 6 through Model Studio and Qwen Cloud.47 Reports from that week agree on the main specifications. The model outputs at 480p, 720p, or 1080p. It doubles the 15-second limit of its predecessor, Wan2.7. Alibaba says it keeps characters, objects, backgrounds, and spatial layout consistent for the full length of a clip.27

The timing drew as much attention as the model. Alibaba priced a share placement of HK$80 billion, about $10.2 billion, the day before the formal launch. Reuters described it as the largest primary follow-on offering ever completed by a Hong Kong-listed company.56 One account also placed the launch a week after Alibaba reported a 75% drop in quarterly earnings, which it attributed to the same AI spending.6 Because the company says all net proceeds from the placement will go to what it calls "full-stack AI," Wan3.0 is one of the first products it can show for that money.410

Thirty seconds changes what counts as a video model

Much of the coverage focuses on the 30-second number, and that is reasonable. Alibaba's own material says most AI video generators produce clips of only a few seconds. It argues that a longer single generation leaves room for pacing, continuous camera moves, and one-take sequences, so the output can carry a short story rather than a single shot.1 Two features support the longer format. A smart-duration mode suggests a clip length based on the prompt, and an extension mode continues an existing generation forward, backward, or in both directions.112

Alibaba describes Wan3.0 as an "All-in-One" model. One endpoint handles text-to-video, first-frame and first-and-last-frame animation, multimodal reference generation, editing, and extension.13 A request can include up to 10 reference images, five reference videos, and five reference audio clips. The reference video and audio are each limited to 15 seconds in total. The output runs at 30 frames per second.412 Text-to-video prompts can be written as multi-shot storyboards, with each shot lasting about four to six seconds and marked with timestamps.13 In practice this turns the model into a basic shot-list tool. That matters more to working editors than any single-frame quality improvement.

Editing, which first appeared in Wan2.7, carries over. Users can change visuals, plot, and dialogue in a generated clip without regenerating it from scratch.1 Alibaba's documentation lists adding, removing, and replacing elements, style conversion, lighting changes, and dialogue edits, all triggered by plain-language instructions in the prompt.12

Documents in, video out

The most unusual feature is document input. Wan3.0 reads DOC, XLS, PPT, PDF, TXT, Markdown, and Apple's Keynote, Pages, and Numbers files. The limit is one file per request, up to 100MB and 50 pages.1 It can also take one public webpage, provided the page does not require a login. A file and a link cannot be combined in the same request.415

Reporters disagree on what this feature is mainly for. Several outlets present it as a route into corporate workflows: turn manuals into training videos, slide decks into product demos, and quarterly reports into short explainers, using material that already sits on company drives.610 Others emphasize the creative uses Alibaba reported during the beta, including short dramas, film production, advertising, tourism promotion, and music videos.56 Our reading is that the enterprise angle is the real strategy. The creative examples show what the model can do, but per-second billing combined with document parsing is built for marketing and communications teams that produce a lot of content, not for one-off artistic projects.7

Sound is now part of the generation

Coverage that only mentions clip length tends to underplay audio. Wan3.0 generates dialogue, background music, and sound effects in the same pass as the picture. Audio is on by default and can be turned off without changing the price.412 Users can write spoken lines and sound cues into the prompt. Alibaba's examples include timed narration and layered sound effects matched to on-screen motion.12 Reference audio can drive lip-sync or dance moves that follow a track's beat or tone.13 Reuters-derived reports also mention synchronized facial micro-expressions and multilingual voice output.27

Some third-party hosting platforms describe the architecture. In their account, video frames pass through a 3D causal VAE and sound through a separate 1D audio VAE, and the two token streams are denoised together in a hybrid MMDiT sequence.14 Alibaba has not published a model card, so that description should be treated as unconfirmed. Alibaba has itself said that audio texture and the accuracy of on-screen text still need work.415 Those are significant admissions for a model built around synchronized sound and turning documents into video.

Images and avatars

The release also covers still images and faces. Alibaba's materials list text-to-image, conversational image editing, generation of up to 12 sequential images in a consistent style, pixel-level box-selection editing, and long-form text rendering in 12 languages, including charts and formulas.3 Portrait controls reportedly reach down to bone structure and eye detail.3 Alibaba's pitch on faces is clear: it promises varied, realistic people instead of the same polished "AI face" repeated across characters, with micro-expressions tied to body language even in group scenes.1

Pricing undercuts Google, and the weights are closed

The list price is $0.05 per second at 480p, $0.10 at 720p, and $0.20 at 1080p.1 A full 30-second clip therefore costs between $1.50 and $6 before discounts.4 Outlets comparing it with Google's Veo 3.1 standard tier, at $0.40 per second, put Wan3.0's top resolution at half that rate.27 Selected platforms also offered a 30% discount on API pricing from August 24 to September 23.8 Alibaba's documentation now lists Wan3.0 as a recommended migration path for developers currently using Sora 2, Veo 3.1, Kling 3.0, Seedance 2.0, and Runway Gen-4. It also offers a faster "prime" variant.17

One point needs care: openness. Earlier Wan releases built a large developer following, and Wan2.2 shipped under Apache 2.0 with downloadable weights.18 An official-looking GitHub repository for Wan3.0 describes the model in open-source terms.3 Other reporting says otherwise. As of August 25, Alibaba's official Wan organizations on GitHub and Hugging Face had no Wan3.0 model files.4 Comparison guides describe it as a closed, API-only model.18 Dataconomy explicitly framed the closed weights as a shift away from the open-source approach that made the Wan series known.7 Our conclusion is that Wan3.0 is a closed, hosted product, and the repository is a showcase rather than a weights release. That fits a company that has just raised $10 billion and now needs revenue from its models. The move toward closed releases did not start here. Wan2.6 had already stopped being open-weight.19

Performance is still unverified

The main gap in the launch coverage is independent benchmarking. Alibaba's announcement included no specific performance metrics.7 Neutral leaderboards had not ranked the model when it was formally released.4 The only structured test in the reporting came from Atlas Cloud, a hosting company that sells access to the models it compared. It ran 49 fixed-seed comparisons against Wan2.7 and ByteDance's Seedance 2.0. It found better identity retention and physical behavior, but also repeated unwanted cuts in shots that were meant to be continuous.4 That flaw undercuts the main promise of a 30-second one-take model, and creators should check for it before paying for 1080p renders.

Access for consumers is also unclear. The formal launch added services aimed at China, including the Qwen app, Qwen Creation for PC, and IF Studio, but eligibility rules were not published.4 The API was still labeled as a preview.4 A consumer site at wan.video was described as members-only.7

Why it matters

Wan3.0 is significant less because of any single feature than because it combines several separate tools in one product. Text-to-video, image animation, reference-based character consistency, voice and sound generation, editing, and document parsing are now one API call, billed by the second. Alibaba has called the eight releases from Wan1.0 to Wan3.0 a move from generating assets to delivering commercial content, with a "world simulator" as the long-term goal.1 That framing is promotional, but the product decisions support the commercial part. Developers lose the ability to run the newest model themselves. In return they get longer clips, built-in audio, and prices well below Google's. For a company spending heavily on AI, that trade-off is deliberate. Whether it pays off depends on benchmarks that had not been published as of this writing.