Alibaba Wan3.0 AI Video Model Adds 30-Second, Document-Driven Clips

What Alibaba shipped

Alibaba has formally launched Wan3.0, the newest model in its Wan video-generation family. The release took place on August 24, 2026, after a public beta that opened on August 6 through Alibaba Cloud's Model Studio and Qwen Cloud.74 It turns text, images, audio, video, documents, and web pages into clips of up to 30 seconds, all in one generation pass.2 Most coverage describes that as twice the 15-second limit of Wan2.7.26 The Decoder is the exception, comparing it to Wan2.5 instead.9 Every other outlet in this round of reporting uses Wan2.7 as the baseline, and Alibaba's API documentation puts that model's text-to-video ceiling at 15 seconds, so Wan2.7 is the right comparison.416

Length gets the headlines, but the bigger change is in how the model interprets its inputs. Wan3.0 marks a move from video models that render whatever they are told toward models expected to read source material, plan a sequence, and decide how long it should run. Alibaba has not described this as reasoning, but that is the real argument for the product.

From prompt-follower to planner

Alibaba's developer blog makes a revealing claim. It says the prompt works as a "control center" that interprets the input, organizes the content, and shapes the final result, and it credits instruction-following for the model's ability to turn very different source materials into coherent video.1 The same post describes a smart duration recommendation feature. The model reads the prompt and suggests a suitable length, so a simple product shot isn't stretched to 30 seconds and an ambitious story isn't cut short.1 Deciding on its own how long a clip should be is a planning task, not a rendering one.

Document input pushes that idea further. Wan3.0 accepts one file or link per request, with a limit of 100MB and 50 pages, in formats including doc, xls, ppt, pdf, txt, Keynote, Pages, Numbers, and Markdown.1 Winbuzzer reports that publicly accessible web pages that don't require a login also count as inputs.4 A slide deck or spreadsheet is not a visual brief. To get from 50 pages of structured information to a coherent 30-second clip, the system has to pick out what matters, put it in order, and present it visually. Pexo's explainer argues that this removes the usual chain of summarizing the deck, storyboarding it, and then generating, and replaces it with a single input.14 Alibaba's own marketing points to training videos built from manuals, product demos made from slide decks, and video versions of quarterly reports.10

The most direct sign of reasoning in the product is a setting most launch coverage overlooked. Fuser's hands-on comparison found that Wan3.0 offers an "Enhanced Reasoning" toggle, off by default, along with an audio-generation switch and a faster Prime tier.17 Fuser ran its tests with that toggle off and did not measure what it does, so its effect has not been shown publicly.17 Still, the existence of the setting tells you how Alibaba thinks about the model. It treats deliberation before generation as an option users can pay attention to, much as language-model vendors now sell thinking modes.

The prompt design points the same way. According to Fuser, Wan3.0 dropped the Multi-Shots toggle that Wan2.6 had. Users now script shots in the prompt itself, with time ranges, or start with "One continuous shot" when they want no cuts. In Fuser's 10-second test, the model cut at exactly 3.0 and 6.0 seconds, as instructed.17 That shifts editing structure from interface controls into natural-language instructions the model has to parse and carry out.

Architecture consolidation

Yotta Labs reports that Wan2.7 launched in April 2026 as four separate models: text-to-video, image-to-video, reference-to-video, and video editing, each with its own endpoint. Wan3.0 is one unified model that covers all of those modes.20 Alibaba's GitHub page lists up to 20 reference assets, instruction-based and reference-based editing, and text rendering in 12 languages, including charts and formulas.3 The API accepts up to 10 reference images, five videos, and five audio clips per request.4 Combining these modes in one model is what makes cross-modal planning possible. Text, pictures, and sound can't be coordinated if they go through separate systems.

Alibaba has also framed the release as part of a longer project. Its blog counts eight iterations from Wan1.0 to Wan3.0 and calls the work progress toward a "world simulator."1 The Decoder notes that Alibaba is promoting the model for producing realistic simulation footage to train autonomous vehicles and robots, as well as for film and marketing.9 That world-model ambition is a reasoning claim too, because the model would need to capture physical cause and effect, not just surface appearance.

Where the reporting diverges

The sources disagree on several points.

Open weights. Alibaba's official GitHub repository describes Wan 3.0 as pushing the limits of "open-source video synthesis."3 Independent reporting contradicts that. Winbuzzer found no Wan3.0 model files on the official Wan organizations on GitHub or Hugging Face as of August 25.4 Fuser confirmed that those organizations host Wan 2.1 and 2.2 but no Wan 2.6 or 3.0 repository.17 Dataconomy calls the closed weights a break from the open-source approach that built the Wan series' reputation.7 OrcaRouter traces how that happened gradually: Wan 2.1 and 2.2 were released under Apache 2.0, Wan 2.5 was announced as open but its weights never appeared, 2.6 was API-only, and 2.7 was only partly open.16 On the evidence, Wan3.0 is a closed, metered product, and the repository's wording is marketing rather than an accurate description.

Audio. Pexo's comparison table lists Wan 3.0 as visual-only with no generated soundtrack.14 That conflicts with Winbuzzer's finding that audio generation is on by default.4 It also conflicts with Fuser's documentation of a Generate Audio switch and Alibaba's statement that the model natively produces dialogue, background music, and sound effects.17 The evidence clearly favors native audio. Pexo, which sells a multi-model video agent, appears to be wrong here.

Release status. Alibaba calls the August 24 launch general availability. However, Fuser notes that the API reference, last updated September 28, still says the model is "currently in preview."17 Winbuzzer also reported that the Model Studio API remained in preview at launch.4

Benchmarks: from silence to the top of the board

At launch, no neutral leaderboard had tested Wan3.0, and Dataconomy pointed out that Alibaba published no specific performance metrics.47 That has since changed. A Pinggy survey from October reports that Wan 3.0 ranks first on Artificial Analysis's text-to-video leaderboard both with audio (1,157) and without (1,128), and sixth on image-to-video with audio.18 Layer's model page lists a different Artificial Analysis Arena Elo of 1362, placing it third of 79 scored models on its platform.12 The difference most likely comes from different boards or snapshots, so readers should treat any single figure with care. The overall direction is consistent, though: within weeks, Wan3.0 moved into the top tier.

Real-world testing has found weaknesses. Atlas Cloud 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. Atlas Cloud sells access to all of those models, so it has a commercial interest in the results.4 For a model whose selling point is understanding instructions, unrequested cuts are a meaningful failure. Alibaba itself acknowledges that audio texture and on-screen text accuracy still need work.4 OrcaRouter adds that longer clips can drift within a single shot, not only at cuts between shots.16

Price and the money behind it

The list price is $0.05 per second at 480p, $0.10 at 720p, and $0.20 at 1080p, so a full 30-second 1080p clip costs $6.9 India Today and Dataconomy compare that with $0.40 per second for Google's Veo 3.1 standard tier.27 OrcaRouter makes a less flattering comparison: at 1080p, Wan3.0 costs 33% more per minute than Wan2.7.16 TechNode reported a temporary 30% discount running through September 23.8 The Decoder notes that a faster Prime tier costs more, at $0.28 per second for 1080p.9

The timing tells you as much as the pricing. Alibaba released Wan3.0 the day after it priced a HK$80 billion (about $10.2 billion) share placement, which outlets describe as the largest primary follow-on offering by a Hong Kong-listed company. All net proceeds are earmarked for AI.619 Technology.org adds that the launch came a week after Alibaba reported a 75% drop in quarterly earnings, driven by AI spending.6

The takeaway

Wan3.0 is less a story about 30-second clips than about video models taking on interpretive work. These tasks include parsing documents, choosing durations, following shot scripts written in plain language, and offering an optional reasoning mode. Alibaba has paired that change with closed weights and per-second billing, which turns what was once an open research line into a revenue product backed by a $10 billion bet. The leaderboard results suggest the bet is technically sound. The unwanted cuts and the undocumented Enhanced Reasoning setting suggest the reasoning claims still need independent testing.