On April 22, 2026, OpenAI dropped Privacy Filter, a 1.5 billion-parameter sparse Mixture-of-Experts model that strips names, addresses, emails, phone numbers, URLs, dates, account numbers and secrets from text before it ever leaves your device. Released under Apache 2.0 on Hugging Face and GitHub, it runs on laptops and even in browsers. At 50 million active parameters, it handles up to 128,000 tokens in one go and nails a 96–97.4 percent F1 score on PII-Masking-300k benchmarks. Startups in the EU face stiff GDPR penalties—up to €20 million or 4 percent of global revenue—if they zap customer support transcripts or feedback into a US-hosted LLM without stripping personal data first. Privacy Filter fills that gap. Instead of brittle regex, it uses bidirectional token classification with BIOES tagging to recognize “Maya Chen who manages the Oslo route” as a single private span. No unredacted text ever touches a remote API. Under the hood, OpenAI converted an autoregressive checkpoint into a token classifier that looks at the whole sequence. It uses grouped-query attention, 128 experts with top-4 routing, banded attention for the long window, and a constrained Viterbi decoder to keep spans coherent. You get 33 output classes—eight PII categories times four BIOES tags, plus one background class. Precision-recall knobs let you tune for high recall in training pipelines or high precision in customer-facing tools. It’s not foolproof. This isn’t true anonymization. Contextual cues in redacted text can still lead to re-identification. The eight categories omit things like SSNs or passports unless you fine-tune. Short, context-poor snippets trip it up. And although it works best in English, other languages need careful evaluation. You still have to test performance on your own data and can’t cite Privacy Filter alone as GDPR compliance.