YecoAI
Model

Yeco LLM Series

Yeco LLM Series is the family of proprietary Italian AI models from YecoAI: Yeco-PII v1.0 (278M, privacy) and Yeco Flash Mini-IT v1 (118M, trained for under €15). Ender-1, our fine-tune of Qwen2.5-Coder-1.5B, is a separate product model. Sovereign, efficient, compliance-ready, this page updates as the series grows.

One model family behind every YecoAI product: privacy, gaming and micro reasoning, all trained in-house. This page grows with the series.

The models at a glance

Yeco-PII v1.0: 278M parameters, neural PII detection and anonymization, native Italian, certified on IT/EN/FR/DE/ES/NL, average F1 0.966 on Ai4Privacy (historical v1.0 protocol; beats OpenAI's Privacy Filter on all 6 languages), about 3,700 chars/s on CPU, one-time license, on-prem delivery. Yeco Flash Mini-IT v1: Italian micro-LLM with 118.1M parameters, custom 32k BPE tokenizer (1.53 tok/word fertility, 34% better than GPT-2), 1.25B pretraining tokens plus SFT on 310K examples, trained for under €15, about 200 tok/s on CPU, Apache 2.0, published with a full model card. Ender-1 is not part of the series: it is a fine-tune of Qwen2.5-Coder-1.5B (Apache 2.0) for EnderDevelopment, in beta.

Yeco-PII v1.0

Our flagship privacy model. 278M parameters that outperform models 5.4x larger on European PII detection, with 24 protected data types and the Italy-specific entities (codice fiscale, VAT number, civil registry data). Read the research paper for the full benchmark methodology, or the product page to start an enterprise API trial.

Ender-1 (Beta)

A fine-tune of Qwen2.5-Coder-1.5B (Apache 2.0), trained on EnderDevelopment usage data, free-tier users under strict policy, now rolling out on the platform that trained it. Ender-1 turns community intent into working code for the gaming stacks communities actually use. It is the first YecoAI model born from a living product and the template for how our data engine works: real usage, governed by policy, compounding quality.

Yeco Flash Mini-IT v1

Proof that sovereignty and efficiency are not opposites: an Italian LLM with 118M parameters trained for under €15, fully documented in a public model card with measured ITA-Bench numbers (BoolQ 62.5 vs Minerva-350M's 60.7 from its paper) and declared limitations. Apache 2.0 license. Checkpoints, tokenizer and training pipeline are reusable starting points for Italian language fine-tuning research.

Related products