Where to use Wan 3.0: wan.video, QwenCloud, or the wan-3.0 API on Sume

There are two kinds of doors to Wan 3.0. Alibaba's own channels, which its README (read 2026-10-05) names as wan.video and QwenCloud, and third-party APIs such as the wan-3.0 model id on Sume. Choose Alibaba's channels if you want Alibaba's own features exactly as Alibaba ships them, including document and webpage references. Choose Sume if you want one API key and one job format for Wan 3.0 next to other video models, plus Sume's media tools for what comes after generation. The model is not open weights: the README and the launch coverage describe a hosted release, so there is no file to download and run yourself.

What the sources say about Wan 3.0

The TechNode report on the launch (read 2026-10-05) is dated 24 August 2026 and describes 30 second generation and document input. The README lists the features: native 30 seconds, up to 20 reference assets (documents, webpages, text, images), instruction- and reference-based video editing, automatic scene splitting, pixel-level box-selection editing, up to 12 sequential images with a cohesive style, and text rendering in 12 languages.

What Sume exposes

Sume's catalog id is wan-3.0. The Video generation docs give its length range as 2 to 30 seconds and say it accepts image, video and audio references. The catalog constraints add: up to 10 images, 5 videos and 5 audio files, and no web_url or file_url field in v1. Video-to-video editing is not listed as a capability for this model, so Alibaba's instruction-based editing and box selection are not request fields on Sume.

Reading this honestly

This post does not compare prices or quality between the channels. I did not read the wan.video or QwenCloud pages in detail, only what the README says about their existence, so any detail about those products should come from Alibaba. For Sume, price is read at call time from pricing_skus in GET /v1/videos/models, which is the only number to plan with.

• If the feature you want is on the README list but not in this table, ask whether it is a request field or a model behavior.
• If you start on Sume and later need a feature Sume does not expose, nothing stops you from using Alibaba's channel for that one job.
• Keep your prompts in files, not in a UI, so they move between doors.

A quick test

Run the same prompt, 5 seconds at 480p, through each door you are considering. Compare the clips, the fields you could set and the time to a usable file. That small experiment is cheaper than a long evaluation, and it shows whether the limits above matter for your work. For Sume the request is one POST /v1/video-router/generate with model: "wan-3.0", a prompt, duration and resolution, and an Idempotency-Key header.