OpenAI Releases Privacy Filter — Its First Open-Weight Model of 2026
OpenAI published Privacy Filter on HuggingFace under Apache 2.0 — a bidirectional token-classification model derived from its GPT-OSS architecture (MoE: 1.5B total / 50M active parameters). The model detects and masks PII in text at scale, supports a 128k context window, can run in-browser on-device, and is positioned for trillion-token data pipeline filtering at minimal inference cost. This marks OpenAI's first open-weight model release of 2026.
Why It Matters
Privacy Filter signals a meaningful shift: OpenAI releasing a practical, Apache-licensed utility model for data processing infrastructure. For teams building data pipelines or RAG systems with sensitive inputs, a 50M active-parameter PII masker from OpenAI at zero licensing cost is immediately deployable. Full model card on HuggingFace.