OpenAI just open-sourced a new model called Privacy Filter under the Apache 2.0 license. It's a locally-deployed text sanitization tool that automatically identifies and redacts eight categories of personally identifiable information (PII) from any text you feed it—names, email addresses, phone numbers, physical addresses, account IDs, URLs, dates, and secret keys. The whole process happens on your own machine, so your sensitive data never has to travel to the cloud.
The model has 1.5 billion parameters total, but it uses a sparse mixture-of-experts architecture. That means only about 50 million parameters are active during any single inference, making it efficient enough to run on a laptop or even in a browser. It can handle a 128K token context window and annotates all the private info in a single forward pass. You can tweak the precision-recall trade-off using preset operating points, or fine-tune it with your own data to adapt to specific use cases. It's primarily an English model, with limited multilingual capabilities.
This local-first approach to privacy is part of a broader trend, as seen with models like Alibaba's Qwen3.6-27B, which also emphasizes running on your own hardware. Meanwhile, OpenAI's other projects, like the recently leaked 'Hermes' agent system, continue to push the boundaries of what AI can automate.