Chinese AI video models by API: Kling, Wan, MiniMax limits and prices

Sume's video API lists four models from Chinese labs that you can call today with one key: kling-3 (Kling, from Beijing-based Kuaishou), wan-3.0 (Alibaba's Wan), and minimax-h3 and minimax-h3-max (MiniMax, based in Shanghai). They share one endpoint, POST /v1/videos, so switching is a change to the model field. The clip lengths, resolutions and reference rules differ, and the table below has them.

What does each model accept on Sume?

Limits come from the Sume video docs, and the catalog endpoint (GET /v1/videos/models) is the live source. Check it before you hard-code a duration or a resolution.

What does a second cost?

Sume bills the provider list price times 1.25. These per-second rates follow from the list prices in the Sume docs and pricing tables, so a 10-second clip is ten times the number shown. Admission reserves the amount first and refunds it if the job fails.

Which one fits which job?

If you are budgeting for a launch you cannot call yet, price it with the closest model you can: a 10-second 1080p clip is $1.40 silent or $2.10 with sound on kling-3, $2.50 on wan-3.0 and $2.00 on minimax-h3-max. Rerun the same numbers when a new id appears in the catalog.

• Longest single shot: wan-3.0 takes up to 30 s in one job; the other three stop at 15 s.
• Cheapest draft: wan-3.0 and both MiniMax models are $0.0625 a second at 480p.
• Silent clip: only kling-3 lets you switch audio off. MiniMax H3 rejects generate_audio false and keeps its sound on.
• Many reference photos, clips or a voice track: wan-3.0, minimax-h3 and minimax-h3-max take audio and video references. kling-3 takes none.
• Fixed 1080p at one price: kling-3 renders 1080p for every request, so a 720p request costs the same.

What about Kling 4.0 and the newest releases?

fal's overview says Kling 4.0 arrives in October 2026 with clips of 3 to 30 seconds, and that it is coming to fal soon (read 2026-10-11). It is not in Sume's documented catalog. Do not assume a model id exists: call GET /v1/videos/models and read the list. Today the Kling id you can submit is kling-3.

Poll the returned polling_url until the job is completed, or send a callback_url over HTTPS and Sume will post a signed webhook when the job finishes. Both paths work the same for all four models.

How do I switch between them in code?

Keep the request body the same and change only the model id, then let the catalog tell you what each model supports. Remember to send an Idempotency-Key so a retry does not bill twice.

curl https://api.sume.com/v1/videos \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Idempotency-Key: demo-0001" \
  -H "Content-Type: application/json" \
  -d '{"model":"wan-3.0","prompt":"A paper boat on a rain-lit street","duration":5,"resolution":"720p","aspect_ratio":"16:9"}'