Models

CoPE is a family of decision models specially trained by Zentropi for online safety use cases. They are the most policy-steerable models of their size and enable accurate, fast, and efficient content classification at scale.

Available models

Scroll horizontally to compare all capabilities

Capabilities of currently available Zentropi models.
Model	Text	Images/video	Context length	Multi-label	System One	Open weights
`cope-b-a4b` — view Hugging Face model card (opens in a new tab) Current text default	Supported	—	256K	—	—	Open weights
`cope-b-a4b-mm` — view Hugging Face model card (opens in a new tab) Current multimodal default	Supported	Supported	256K	—	—	Subscriber-only
`cope-b-s1` — view Hugging Face model card (opens in a new tab) Multi-label & native System One	Supported	—	256K	Supported, up to 26 labelersUp to 26	Native System One format supported	Open weights

Our first generation model, cope-a-9b, will be deprecated in our API at the end of 2026. Model weights will continue to be available.

Choosing a model

Text: Choose cope-b-a4b for efficient and accurate text evaluation. The cope-latest alias resolves to cope-b-a4b for text.
Images or video: Use cope-b-a4b-mm. All hosted usage of this model is subscriber-only. The cope-latest alias resolves to cope-b-a4b-mm for images or video.
Multi-label: Explicitly select cope-b-s1 for /v1/multilabel, with up to 26 labelers per request.
System One: Self-host cope-b-s1 if you need full decision model support, including noul, choice, and score request types.