Wan 3.0 is a closed beta and API, not open weights: a Sume call

Wan 3.0 is not an open-weights release. The Alibaba README (read 2026-10-05) points to wan.video and QwenCloud and describes a closed beta and API. If you searched "Wan 3.0 open source" or "Wan 3.0 Hugging Face", the answer from the vendor page is that there is no downloadable model there. On Sume, wan-3.0 is a catalog id you call over HTTP, with a 2 to 30 second range.

What the README lists

TechNode reported the launch on 2026-08-24 (read 2026-10-05). Use the vendor README for feature claims and check what a hosted route accepts separately.

• Native 30-second generation.
• Up to 20 reference assets: documents, webpages, text and images.
• Instruction- and reference-based video editing, plus pixel-level box-selection editing.
• Auto scene splitting and up to 12 sequential images with a cohesive style.
• Text rendering in 12 languages.

What a Sume call looks like

Sume's POST /v1/videos agrees field for field with the OpenRouter video generation API, so a client written for that wire works after you change the base URL and key. The required fields are model and prompt. Sume's docs list wan-3.0 at 2 to 30 seconds, and say Wan 3.0 accepts audio and video references as well as images.

curl -X POST https://api.sume.com/v1/videos \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: wan-3-001" \
  -d '{
    "model": "wan-3.0",
    "prompt": "A paper boat drifts down a rain gutter, close tracking shot",
    "duration": 15,
    "resolution": "720p"
  }'

Moving from the vendor's route to Sume

If you already tested Wan 3.0 on the vendor's site, your prompts carry over, but your settings may not. Re-check duration, resolution and aspect ratio against the Sume catalog entry, as the hosted route validates them. Send an Idempotency-Key on every create so a retry does not queue a second job, and poll the job envelope for the result.

Keep your prompts in files so you can compare outputs from both routes. If results differ, note the settings first; differences in resolution and duration explain most of it.

• Re-check duration and resolution against the catalog.
• Always send an idempotency key.
• Keep prompts in files, not chats.

Polling and results

The create call returns an async job. Poll GET /v1/jobs/:id/status or use a webhook, and fetch the result when it is ready. Keep the idempotency key stable across retries of the same request, and use a new one for a new take. If a call returns a validation error, read the code and compare the parameter to the catalog entry's supported values. Common causes are a duration outside 2 to 30 seconds for this id, a resolution the model does not list, or a reference type its supported_input_references does not include. Fix the field and send the request again under a new idempotency key.

What the hosted route does not show you

"Not listed" means the Sume pages I read do not describe the feature. Read the model's entry in GET /v1/videos/models before you promise a client any of them.