CoPE

CoPE (Content Policy Evaluation) is a small open model from Zentropi trained to do one thing: read a policy you wrote and decide whether a piece of content violates it. Instead of shipping a fixed taxonomy, it takes the policy as part of the input, so a marketplace and a dating app can run the same model with different rules and get different answers.

At around 4B active parameters it runs on a single modest GPU. It's aimed at teams who want to write policy in plain language and get consistent, explainable labels back, and Zentropi also offers a hosted version with a policy authoring workspace.