OpenAI Privacy Filter
OpenAI's Apache 2.0 token classifier that finds and masks personal data in text: 1.5B parameters, 50M active, 128K context.
In short. OpenAI Privacy Filter is a small model that reads text and marks personal data, such as names, emails, phone numbers and passwords, so you can mask it before the text goes anywhere else. It does not write text. It runs on a laptop or even in a web browser, and it uses the Apache 2.0 license.
Summary
OpenAI announced Privacy Filter on 22 April 2026 as an open-weight model for detecting and redacting personally identifiable information (PII) in text. It is not a chat model. It starts from an autoregressive checkpoint with an architecture similar to gpt-oss but smaller, converted into a bidirectional token classifier: an 8-layer stack with a residual width of 640, grouped-query attention over a banded window of 257 tokens, and sparse mixture-of-experts layers with 128 experts and 4 used per token. The Hugging Face file holds about 1.4 billion parameters, and OpenAI states 1.5 billion in total and 50 million active. The model labels every token in one pass and then decodes whole spans with a constrained Viterbi step. The context window is 128K tokens, so long documents need no chunking.
It detects eight categories: account numbers, private addresses, emails, person names, phone numbers, URLs, dates and secrets such as passwords and keys. Runtime settings let you trade precision against recall and change how long the detected spans are. OpenAI reports a token-level F1 of 0.960 on the PII-Masking-300k test split, and 0.974 on a version whose annotation problems OpenAI corrected. On the CredData secrets benchmark its token-level F1 is 0.844 but its span-level F1 is only 0.624, because exact span boundaries differ. OpenAI also reports that fine-tuning on a small amount of data lifted F1 on a domain benchmark from 54 percent to 96 percent. All of these figures are OpenAI's own, and OpenAI says it uses a fine-tuned version of the model in its own privacy workflows.
The BF16 file is 2.8 GB, and the repository also holds ONNX files down to a 4-bit build of about 0.9 GB that runs in a browser through Transformers.js with WebGPU. The model card stresses that it is a redaction and data-minimization aid, not an anonymization or compliance guarantee. The label set is fixed, so a different policy needs fine-tuning, and the card warns that it can miss unusual names, non-Latin scripts or new credential formats, and can over-mask public names or harmless high-entropy strings. In the 30 days to 11 November 2026, Hugging Face counted about 103,000 downloads.
Good for
- Masking names, emails, phone numbers and secrets in logs, tickets and documents before sharing, indexing or training on them
- Running redaction locally or in a browser so the raw text never leaves the machine
- Fine-tuning a small classifier to an organization's own privacy policy
- A fast first pass in front of a human review step in sensitive workflows
Watch out for
- OpenAI calls it a data-minimization aid, not an anonymization or compliance guarantee, and asks for human review in medical, legal, financial and HR settings
- The label set is fixed at eight categories, and a different policy needs fine-tuning
- Accuracy can drop on non-English text, non-Latin scripts and unusual naming patterns, and span boundaries can shift, most clearly for secrets in code
- All scores are OpenAI's own, and the "corrected" benchmark uses OpenAI's own revision of the dataset labels
- It only labels text; it cannot rewrite or answer questions
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Estimates for the Privacy Filter (BF16) variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.
| Variants | Parameters | BF16 | FP8 / INT8 | 4-bit (Q4) | Fits on |
|---|---|---|---|---|---|
| Privacy Filter (BF16) BF16 | 1.4B 50M active | 3.22 GB | 1.61 GB | 0.966 GB | Laptop or phone |
| ONNX q4 (Transformers.js with WebGPU) INT4 | 1.4B 50M active | 3.22 GB | 1.61 GB | 0.99 GB | Laptop or phone |
Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software.
The license in plain words
One of the most widely used permissive licenses. You may use, change and sell products built on the model with only light conditions.
Frequently asked questions
- What is OpenAI Privacy Filter?
- OpenAI Privacy Filter is a small model that reads text and marks personal data, such as names, emails, phone numbers and passwords, so you can mask it before the text goes anywhere else. It does not write text. It runs on a laptop or even in a web browser, and it uses the Apache 2.0 license.
- Can I use OpenAI Privacy Filter commercially?
- Yes. OpenAI Privacy Filter is released under Apache 2.0, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
- Is OpenAI Privacy Filter open source?
- OpenAI Privacy Filter is released under Apache 2.0, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
- Can I run OpenAI Privacy Filter on my own computer?
- Yes. The smallest variant, Privacy Filter (BF16), needs about 0.966 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
- When was OpenAI Privacy Filter released?
- OpenAI Privacy Filter was released by OpenAI on April 22, 2026.
- How long a context window does OpenAI Privacy Filter have?
- OpenAI Privacy Filter supports a context window of up to 128K tokens.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- Introducing OpenAI Privacy Filter (OpenAI, 22 April 2026) openai.com
- Model card for OpenAI Privacy Filter (OpenAI, PDF) cdn.openai.com
- openai/privacy-filter model card (Hugging Face) huggingface.co
- openai/privacy-filter repository (GitHub) github.com
Release date and facts last verified November 11, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| gpt-oss OpenAI | 21B – 117B 5.1B active | 128K | Apache 2.0 (Permissive) | 15 GB Laptop or phone | |
| Clef Cloudflare | 9.41B – 27.4B | 64K | Apache 2.0 (Permissive) | 6.49 GB Laptop or phone | |
| Liquid d1 Liquid AI | 587M – 3.12B | 128K | LFM Open 1.0 (Conditional) | 0.405 GB Laptop or phone | |
| Needle 3 Cactus Compute | 121M | 8K | Apache 2.0 (Permissive) | 0.0835 GB Laptop or phone |