Yeco Flash Mini-IT v1
118M-parameter Italian micro-LLM, trained for under €15, ~200 tok/s on CPU. Beats 5× larger models on Italian logic benchmarks.
Paper
Yeco-PII v1.0
Proprietary 278M model. Beats OpenAI Privacy Filter 1.5B (F1 0.966 vs 0.769) on all 6 languages.
Abstract
We present Yeco-PII v1.0, a 278M-parameter neural model for PII detection and anonymization that beats the OpenAI Privacy Filter (1.5B) with 5.4× fewer parameters: F1 0.966 vs 0.769, winning all 6 European languages.
Benchmarks
Evaluated on real-world rows never seen in training, with third-party annotations (Ai4Privacy). Same test for every model. Gaps of 20 to 38 F1 points in our favor on every language.
- 278M parameters · ~1.1 GB, standard HuggingFace format
- 24 protected entity types: names, tax IDs, VAT, IBAN, cards, addresses, cadastral data
- Certified on IT, EN, FR, DE, ES, NL — native Italian
- ~3,700 characters/second on CPU · ~1 GB RAM at scale
Why size is not capability
Detection is not a brute-force problem. Checksum networks (mod-97 for IBAN, Luhn for cards) only catch what a formula can validate. Names, addresses, organizations and dates have no checksum: real contextual learning is required, and that is exactly where larger competitors lose 20-30 F1 points.
Conclusion
Compact, rigorously evaluated models can out-engineer giants on vertical tasks. Yeco-PII is proprietary software; full technical documentation is reserved for licensed clients.