Multimodal input

Multimodal workloads pass images, audio, video, or PDFs in addition to text. The catalogue treats modality as a typed list — each verified entry is one of text-in, image-in, audio-in, video-in, text-out (and the matching output channels) — and refuses to assert a channel the vendor docs do not list.

That is the safest stance the catalogue can take, but it has two consequences readers should hold in mind:

• Absence of a channel is not a rejection. A model may informally accept an image without the vendor listing image-in on its model card. The shortlist below filters on the verified field only — confirm against the vendor docs before excluding a candidate. Some Sprint-18 records (Mistral Large 3, the Llama 4 family) carry an intentionally null modality list because the vendor card describes the model as "multimodal" without enumerating the channels.
• Output channels matter too. Most chat models are text-out only — image-generation, audio, and video outputs are modeled separately. The catalogue records output channels on the same modality field; reading both ends gives the full picture of what the model can produce.

For most workloads the safest progression is: filter on the specific input channel you need; open the model record; confirm against the vendor source URL; only then move to /compare with a verified peer.