LangWatch
Sign Up

OpenAIOpen weights model1.4B parameters

OpenAI Privacy Filter benchmarks

All models
  • General rank
    Too few tasks3 non-coding tasks ranked5 needed for a rank
  • Cost
    Self-hostedSmall GPU, ~3.3 GBNo hosted price
  • Multilingual#16 / 43 tied
    31.8%Best task · PII left invs PII detectors

Comparison summary

Strongest
Safety and guardrails
Leads 0 of 3 (0%)

An entity tagger: it marks spans of text such as names, addresses or account numbers, used to find personal data and mask it. Strongest at safety.

Specifications

Lab
OpenAI
Weights
Open weights
Parameters
1.4B
Licence
Apache-2.0
Type
encoder-NER

Safety and guardrails

Share of tasks led · #19 of 46

  1. Gemma 4 31BLeads 3 of 4
  2. Perplexity PII TracerLeads 2 of 4
  3. Amazon ComprehendLeads 1 of 4
  4. Qwen3.8 27BLeads 1 of 5
  5. Gemma 4 12BLeads 0 of 4
  6. Ministral 3 14BLeads 0 of 5
  7. Gemma 4 26B A4BLeads 0 of 4
  8. OpenAI Privacy Filter
    Leads 0 of 3
See the 5 tasks

Traces

Tier 7 of 19#17 / 46

PII left in, % · Lower is better

  1. Amazon Comprehend10.6, interval 9.4 to 11.9, ahead alone
  2. GLiNER Multi PII v115.3, interval 13.2 to 17.3, tier 2
  3. NVIDIA GLiNER PII19.7, interval 17.4 to 22.0, tier 3
  4. Gemma 4 26B A4B19.8, interval 15.8 to 24.2, tier 3
  5. Qwen3.8 27B20.3, interval 16.5 to 24.3, tier 3
  6. GLiNER2 Large20.4, interval 18.0 to 22.8, tier 3
  7. Gemma 4 12B23.3, interval 19.0 to 27.6, tier 3
  8. OpenAI Privacy Filter
    44.8, interval 40.5 to 49.1, tier 7
Tested against 52 PII detectors

Business

Tier 6 of 18#18 / 46

PII left in, % · Lower is better

  1. Perplexity PII Tracer2.1, interval 1.2 to 3.0, tied for the lead
  2. Qwen3.8 27B2.4, interval 1.7 to 3.3, tied for the lead
  3. Gemma 4 31B2.9, interval 1.9 to 4.0, tied for the lead
  4. GLiNER Multi PII v14.1, interval 3.1 to 5.3, tier 2
  5. Ministral 3 14B5.4, interval 4.0 to 6.9, tier 2
  6. Gemma 4 12B5.8, interval 4.4 to 7.3, tier 3
  7. Amazon Comprehend5.9, interval 4.6 to 7.2, tier 3
  8. OpenAI Privacy Filter
    16.3, interval 14.1 to 18.7, tier 6
Tested against 53 PII detectors

Multilingual

Tier 11 of 25#16 / 43

PII left in, % · Lower is better

  1. Perplexity PII Tracer2.9, interval 2.5 to 3.4, ahead alone
  2. Qwen3.8 27B4.2, interval 3.7 to 4.8, tier 2
  3. Gemma 4 31B4.5, interval 3.9 to 5.1, tier 2
  4. GLiNER Multi PII v15.7, interval 5.2 to 6.3, tier 3
  5. bardsai EU PII Multilang6.9, interval 6.3 to 7.5, tier 4
  6. Qwen3.6 35B A3B7.7, interval 7.0 to 8.4, tier 4
  7. Gemma 4 12B8.9, interval 8.1 to 9.7, tier 5
  8. OpenAI Privacy Filter
    31.8, interval 30.2 to 33.4, tier 11
Tested against 53 PII detectors

Look-alikes

Not ranked

Look-alikes redacted, % · Lower is better

  1. Gemma 4 31B3.5, interval 2.1 to 5.1, ahead alone
  2. Qwen3.8 27B10.3, interval 7.9 to 12.9, tier 2
  3. NuExtract 319.8, interval 15.9 to 23.6, tier 3
  4. Qwen3.5 9B20.5, interval 16.8 to 24.3, tier 3
  5. Ministral 3 14B30.5, interval 26.5 to 34.8, tier 4
  6. NVIDIA GLiNER PII42.0, interval 38.2 to 46.1, tier 5
  7. Gemma 3 4B IT45.0, interval 41.0 to 49.0, tier 5
  8. OpenAI Privacy Filter
    49.4, interval 44.9 to 53.9, not ranked
Tested against 53 PII detectors

LangWatch policy

Not ranked

PII left in, % · Lower is better

  1. Gemma 4 31B6.9, interval 2.8 to 11.3, ahead alone
  2. Ministral 3 14B10.2, interval 5.8 to 15.4, tier 2
  3. OpenAI Privacy Filter
    13.0, interval 9.0 to 17.4, not ranked
  4. Gemma 4 12B16.5, interval 10.3 to 23.2, tier 3
  5. Qwen3.8 27B16.7, interval 9.8 to 25.0, tier 3
  6. Gemma 4 26B A4B20.5, interval 13.3 to 28.8, tier 4
  7. Qwen3.5 9B21.5, interval 13.4 to 30.6, tier 4
  8. Gemma 3 4B IT22.4, interval 14.3 to 32.1, tier 4
Tested against 57 PII detectors

Cost

Self-hosted · GPU memory needed, among encoder-NERs · lower is better

  1. Gravitee BERT Small PII~0.6 GB
  2. Secret Masker v3.3a~0.8 GB
  3. Anonym-IA CamemBERT PII FR~0.9 GB
  4. BERT base NER (dslim)~0.9 GB
  5. Veil PII KO Lite~0.9 GB
  6. bardsai EU PII Multilang~1.6 GB
  7. XLM-RoBERTa NER HRL~1.6 GB
  8. OpenAI Privacy Filter
    ~3.3 GB
How we count

A model leads a task when it is in the task's leading tie tier. Each task counts inside its own benchmark, against that benchmark's models; no score is averaged. An area chart shows the models ranked on at least 2 of the area's tasks in a benchmark this model is in too. The overall standing counts every non-coding tasks and needs 5 for a rank. Latency is compared only on one hardware tier.

OpenAI Privacy FilterOther modelsWhisker: 95% interval