Skip to content

Multimodal Models

Curated capability table for judge models used in multimodal evaluation.

This is the first ("blessed list") layer of judge-model capability resolution. It maps the judge models people actually configure to whether they accept image input, so the common case is answered instantly with no network call. Models not covered here fall through to the explicit skip_multimodal_check flag and then to an empirical probe.

The list is intentionally small and conservative: it should only contain model families we are confident about. It does not need to be exhaustive, because the probe covers the long tail.

lookup_blessed_multimodal(model_id)

Return the curated image-support verdict for a judge model id.

Parameters:

Name Type Description Default
model_id str | None

The judge model id, typically provider/model.

required

Returns:

Type Description
bool | None

bool | None: True if the model is known to accept images, False if it is known not to, or None when the model is not in the list and capability must be resolved another way (caller flag or probe).