New AI video models since June 2026: which run on Sume?

Which of the new video models run on Sume? Seedance 2.5, Wan 3.0, MiniMax H3 and Gemini Omni 1.1 Flash do, each under its own id in the video router. Luma Ray3.2 and Kling 4.0 do not. The table below gives one fact from each vendor's own page and the Sume id, if any.

Reading the table

Vendor facts come from Luma's Ray3.2 post, the Wan 3.0 repository, the MiniMax guide, the Omni docs, the Seed announcement and the Kling 4.0 post. Sume ids come from the catalog in the video models docs.

Sume's own durations differ from some vendor pages. Wan 3.0 runs 2 to 30 seconds there, MiniMax H3 starts at 5, and Omni runs 3 to 10 per clip. Sume does not expose Omni's 40 second extension.

Why two are missing

Luma's post says its API is available for the first time with Ray3.2, but the page lists no pricing, and Sume has no Luma row. Kling 4.0 is in early access with no API details on the page. Sume does not list a model until it has a catalog row it can route and bill, so a launch announcement on its own changes nothing in the list.

Veo is separate: Sume's docs say Veo 3.1 is not in the catalog.

Check it yourself

The authoritative list is GET /v1/videos/models. One quirk: higgsfield-genjutsu appears there only when its provider is configured, so your account may list 11 ids where the docs table shows 12.

What to do with this

Treat the table as a snapshot dated 2026-10-06. If your plan depends on a model that is not on Sume, pick the nearest catalog row for the same job and test it. For long clips, Seedance 2.5 and Wan 3.0 are the two rows that reach 30 seconds. For short clips with native audio, Omni covers 3 to 10 seconds. For multi-reference work, the MiniMax rows take up to 9 images, 3 videos and 3 audio clips, with 12 files combined.

Whenever a vendor post says an API is coming, wait for a stable model id and a published price before you plan a launch around it.