Model Card｜Wan3.0

TL;DR Wan3.0: single-shot length doubles from Wan2.7's 15s to 30s, up to 1080P, supports doc/xls/ppt/pdf/md files and web pages as generation inputs, priced at 480P $0.05 / 720P $0.10 / 1080P $0.20 per second — roughly 50% cheaper than Google Veo 3.1 Standard, but now closed-source API-only, and not yet independently tested by third parties

#ai-agent #model-release #daily #alibaba #model-family-wan

Table of Contents

Model Information
Key Capabilities
Benchmark Performance
Comparison with Predecessors and Competitors
Implications for Agent Development
Takeaway
References

Series: AI Model Tracker (6 / 10)

Model Card｜GLM-5.3 Model Card｜Qwen3.8-Flash-Next

🌏 中文版

Model Information

Field	Value
Model ID	`wan3.0-video`
Vendor	Alibaba Cloud (Tongyi Wan team)
Parameters	Undisclosed
Context Window	Not a text model — no context window concept; max single-shot generation is 30 seconds at up to 1080P/30fps
Input Pricing (USD/1M tokens)	N/A (billed per generated second; see pricing below)
Output Pricing (USD/1M tokens)	480P $0.05/sec, 720P $0.10/sec, 1080P $0.20/sec (a 30-sec 1080P video ≈ $6.00)
Open Source	No (Wan 2.1/2.2 were open-sourced under Apache 2.0; Wan3.0 onward is API-only, closed-source)
Release Date	2026-08-24 (public beta since 2026-08-06)
Official Announcement	Alibaba Cloud Model Studio: Wan 3.0
HuggingFace	None (closed-source, not listed)
Family	Wan 3.x (predecessor: Wan2.7, released 2026-04-03)

Key Capabilities

Single-shot generation length doubles from Wan2.7's 15 seconds to 30 seconds — one of the few mainstream video models that supports half-minute single-take narratives
New "Omni Reference": doc, xls, ppt, pdf, md files and web pages can all serve as generation inputs, converting tables, slides, and document content directly into video
Up to 10 reference images can simultaneously lock character, scene, and prop consistency, maintaining appearance and spatial relationships across multi-shot transitions
Native audio-visual synchronized generation with multilingual lip-sync (official demo shows a single character singing consecutively in 8 languages)

Benchmark Performance

Comparison	Wan3.0	Wan2.7 (predecessor)	Strongest competitor (Google Veo 3.1 Standard)
Single-shot length	30 sec	15 sec	No clear public limit (Veo 3.1 ≈ 8 sec per shot)
Max resolution	1080P	1080P	1080P
1080P pricing (per second)	$0.20	Not listed per-second (billed per generation)	$0.40
Document/web page input	Supported (doc/xls/ppt/pdf/md/web pages)	Not supported	Not supported

⚠️ The generation length, consistency, and lip-sync quality claims above are all from Alibaba's official demos. No third-party independent testing or standardized benchmark scores exist yet; pricing figures are from official announcements.

Comparison with Predecessors and Competitors

Compared to Wan2.7, the biggest improvement is a straight doubling of generation length (15s → 30s), plus the new Omni Reference capability that lets any file type serve as input. Previously you could only feed images, video, and audio — now slides, spreadsheets, PDFs, and web pages can all become video script inputs. This is a meaningful product differentiation for marketing, education, and enterprise content production, rather than just parameter stacking.

Compared to Google Veo 3.1, Wan3.0 at $0.20 per second for 1080P is half the price of Veo 3.1 Standard's $0.40, a clear cost advantage. However, Veo 3.1 has more third-party testing and ecosystem integrations. Wan3.0 is currently only available through an application-based API preview and has not fully opened access — this has been repeatedly noted in coverage as a caveat that "quality claims remain to be verified."

A notable strategic shift: Wan 2.1/2.2 were among the few mainstream video models to open-source weights on Hugging Face/GitHub (Apache 2.0). From Wan 2.5 onward, the series has progressively moved to API-only, and Wan3.0 continues this closed-source trajectory. If you previously relied on Wan's open-source versions for self-hosted inference, Wan 2.2 is currently the last downloadable version — Wan3.0 is only available through the cloud API.

Implications for Agent Development

Omni Reference turns "document-to-video" into a single API call, directly impacting agents that build content automation pipelines. Previously you'd need a separate agent to summarize a PDF/slide deck into a text script before feeding it to a video model — now you can skip the intermediate summarization step and pass the raw document directly to Wan3.0.

If you're building marketing/educational content automation agents: you can design a single-step workflow of "upload slides → directly produce a 30-second short video," without maintaining a document-to-script middleware layer
If you're building multimodal content review or localization agents: the 30-second length plus multilingual lip-sync is suitable for batch-generating multiple language versions of the same video
Not suitable for: scenarios requiring offline/local deployment (Wan3.0 has no open-source weights, cloud API only), and production content requiring strict quality guarantees (quality claims are not yet third-party verified, and the API is still in application-based preview — stability and quotas remain to be seen)

Takeaway

I used to think the "video model arms race" would only push toward higher resolution and longer generation length. But Wan3.0's Omni Reference (accepting doc/xls/ppt/pdf/md as generation inputs) is a reminder that real product differentiation often lies in "how open the input interface is" rather than "how impressive the output specs are." It's also a reminder that open source is not a permanent vendor commitment — Wan went from Apache 2.0 at 2.1/2.2 all the way to fully API-only at 3.0, which means when evaluating models you can't just check "has this family ever been open source" — you need to look at the actual license terms of the latest version.