Zentropi says the CoPE model can remove bottlenecks in large-scale content moderation

Zentropi co-founder Dave Willner says its open-source CoPE (Content Policy Evaluator) model reads policy text directly at inference time, turning content standard updates from retraining models into editing documents. CoPE has 9 billion parameters, and its ability to label content according to arbitrary standards is comparable to GPT-4o, with roughly 1% of the parameter count and latency below 200ms on a single consumer GPU. The author says using real samples from X, they generated a labelers in about an hour to block requests that undress or sexualize real people, but it has not yet been tested on a second dataset and is expected to have slightly lower accuracy on other samples.