On 24 August 2026, Alibaba widened access to Wan 3.0, its video model developed by Tongyi Lab, in public beta since 6 August. Two features stand out: thirty seconds of sound video in a single pass, double the previous version, and above all the ability to accept documents as input. A spreadsheet, a presentation or a PDF can become a video, an angle that direct competitors do not occupy.
At the end of July we put together a comparison of video models, concluding that none won on every axis. This release confirms that diagnosis while adding an axis nobody was thinking about.
What Wan 3.0 brings
Thirty seconds in a single pass. That is double Wan 2.7, and it puts the model on a par with the best in the category, reached before it by Seedance 2.5. The difference from stitching together short shots is not cosmetic: a continuous shot allows uninterrupted camera movement and sequence-shot writing, impossible to achieve by splicing pieces together.
Sound at the same time as image. The video comes back with audio rather than silent, in the same generation pass.
Consistency over time. Alibaba claims to have tackled the visual drift and facial deformation that affect long clips, with up to twenty references kept consistent across the whole shot.
A duration recommendation, with the system suggesting an optimal length based on the request, and an extension capability to prolong a narrative.
Document input, a rare capability
This is the point that most distinguishes this release, and it went unnoticed in most coverage.
In addition to text, image, audio and video, Wan 3.0 accepts office documents: word processing files, spreadsheets, presentations, PDFs, as well as web pages. The model reads them and builds a video from their content.
Alibaba presents this capability as an industry first. We found no independent source to confirm that, so we stick to a more modest observation: the competing models we examined announce four input modalities — text, image, audio and video — without mentioning office documents. So it is at minimum a sparsely occupied axis.
Concretely, this means a sales presentation, a quarterly report or a product sheet can become a video without going through the step of writing a script. It is a very practical professional use case, and it targets a market the other models do not directly address: companies that already produce huge amounts of documents and would like to repurpose them as video.
Where it stands against the others
| Model | Duration in one pass | Indicative pricing | Strong point |
|---|---|---|---|
| Wan 3.0 (Alibaba) | 30 s | 0.3 to 1.2 yuan/second depending on resolution | Document input, a first |
| Seedance 2.5 (ByteDance) | 30 s | Varies by platform | 4K, up to 50 references |
| FLUX 3 (Black Forest Labs) | 20 s | Not disclosed | Unified architecture, robotics ambition |
| Kling 3.0 (Kuaishou) | 15 s | ~$0.11 to $0.14/s (Turbo) | Native 4K at 60 fps, per-character lip-sync |
| Gemini Omni Flash (Google) | ~10 s (capped preview) | ~$0.10/s | Top of blind rankings |
| MiniMax H3 | short shots, ~30 s only via extension | Varies by platform | Widely available, shot extension |
Two readings emerge, and the first needs a clarification. On duration, Wan 3.0 does not set a record: Seedance 2.5 was already producing thirty seconds in a single pass since June. Alibaba is therefore matching the top tier, not moving it. And the MiniMax H3 line deserves a careful read: reaching thirty seconds through successive extension is not the same thing, since every cut is visible and rules out continuous camera movement. On price, however, the positioning is aggressive.
But the most important caveat lies elsewhere: none of these results has been measured independently. The quality claims rest on Alibaba's statements. As we remind readers in our article on benchmarks, an announced capability is not a verified capability, and a long duration says nothing about the quality of what happens during those thirty seconds.
This is the point worth highlighting, because it is telling. Wan 3.0 had been pre-announced as a release with open weights under a permissive licence. It ultimately ships in controlled access, by application, with no downloadable weights. The open weights in the Wan series stop at version 2.2. The community had in fact greeted the initial promise with scepticism rather than enthusiasm, given a track record: other versions announced as open never were. That scepticism was justified.
The financial context, which is not incidental
The timing is worth noting. This access expansion comes the day after a share placement in Hong Kong that raised around 80 billion Hong Kong dollars, nearly US$10.2 billion, described as the largest operation of its kind ever carried out by a company listed on that exchange. The entire net proceeds are earmarked for artificial intelligence.
Alibaba has also reported 45% year-on-year growth in its cloud and AI revenue for the June quarter.
The shift from an openness strategy to a closed model therefore fits a deliberate monetisation logic. This is not inconsistent with what we observe elsewhere: we have described the publication of open language models by the same group. Alibaba opens on text and closes on video, which suggests video is seen as the segment where value is captured directly.
What this changes for a creator
Three practical consequences.
The duration barrier falls. Thirty seconds in a continuous shot covers a full advertising format or an entire narrative sequence. It is a threshold that changes what can be envisaged without editing.
Access remains constrained. The model goes through an application process on Chinese platforms, with a consumer offering announced by invitation. For a European creator, practical availability is far from guaranteed.
The choice remains a question of axis. That is the conclusion we drew in July and it holds: no model wins everywhere. For a single long take, Wan 3.0 or Seedance 2.5. For technical finish in 4K, Kling. For the best value for money, Gemini Omni Flash. For immediate availability without a regional account, MiniMax H3. And for starting from existing documents, Wan 3.0 is today the most direct candidate.
What to take away
The race for duration is probably reaching a useful plateau: beyond thirty seconds, editing becomes the right answer again, and the real challenge shifts to consistency and art direction.
What makes this release interesting, then, is not the duration, already achieved by a competitor. It is the idea of document input, which shifts the question from how do I describe what I want to what do I do with what I already have. Many companies do not need a tool that invents: they need a tool that repurposes. That is a commercial angle the Western models, focused on creative prowess, do not occupy.
Then there is the underlying issue, and it extends beyond Alibaba: when a promise of public openness turns into application-based access, it is a reminder that in this sector, openness is a business strategy rather than a value. It lasts as long as it serves, and not one version longer.