OpenAI's Privacy Filter vs GLiNER on 600 PII samples

Both models are open weight, both run on a local CPU workstation, both detect PII in text. Quick rundown of what I found. GLiNER large-v2.1 is ~300M params, zero shot, you pass entity types as plain text strings at inference. Openai/privacy-filter is 1.5B total but only 50M active per forward pass thanks to a sparse MoE. In practice on CPU openai/privacy-filter ran ~2.8 samples/sec vs ~1.1 for GLiNER large. Eval was 400 English + 200 multilingual samples from ai4privacy/pii-masking-400k, s