OpenAI released Privacy Filter, a 1.5B-parameter Apache 2.0 model that runs locally to detect and mask names, addresses, emails, dates and other PII before text ever reaches a cloud API. It offers high benchmark performance, a 128K token context window, and configurable precision-recall, helping startups meet GDPR and similar data-transfer rules. The article covers its architecture, real-world use cases, tuning steps and limitations around true anonymization.

• OpenAI released "Privacy Filter," a 1.5B-parameter (50M active) MoE model under Apache 2.0 that locally masks PII before text hits a cloud API, hitting 96-97.4% F1 on PII-Masking-300k with a 128K context window.
• It's aimed at helping startups avoid GDPR fines (up to €20M or 4% of global revenue) when sending user data to US-hosted LLMs, and uses bidirectional token classification (BIOES tagging) instead of regex to catch contextual PII spans.
• It's explicitly not true anonymization—re-identification from context is still possible, only 8 PII categories are covered by default (no SSNs/passports without fine-tuning), and it performs best only in English.