Hy Image 3.5 reference limit: 5 or 20, and Sume's caps

The Hy Image 3.5 preview accepts 5 reference images in the announcement but up to 20 in the API guide, according to CellCog's write-up. When sources disagree, test the limit yourself. On Sume, the reference cap is a per-model value you read from the catalog, and the documented examples range from 5 to 16.

What the preview post says

CellCog's preview post reports a September 22, 2026 preview, $0.024 per image internationally, up to 4096 by 4096 output (other documentation conflicts and says 2K), a 100k-token prompt, a proprietary model with no weights, and the 5 versus 20 reference conflict.

Reference caps on Sume

Sume's Image API publishes capability descriptors per model, so you can read the real limit before you call. In the docs: ChatGPT Image 2.5 supports up to 16 image references, Ideogram 4.5 uses the first image for editing plus up to 4 more (5 total), and the sample descriptor for Seedream 4.5 shows input_references from 0 to 10. A model whose descriptor is {"min": 0, "max": 0} is text-to-image only and rejects references.

I did not find a Hy Image id in the Sume docs I read, so check GET /v1/images/models for current availability.

How to settle a conflicting limit

• Send a request with the larger number and read the error message.
• Record the limit and the date you tested it.
• Design your product for the lower number so a change upstream cannot break it.
• On Sume, read supported_parameters.input_references instead of testing.