YecoAI

Yeco Flash Mini-IT: an Italian Micro-LLM Trained for Under €15

118M parameters, trained for under €15, ~200 tok/s on a consumer CPU. Yeco Flash Mini-IT is our proof of concept that sovereignty beats scale.

Marco Nasi

Founder & CEO

Frontier AI labs spend millions on training runs. YecoAI took the opposite path: prove how far a disciplined, fully documented pipeline can go on a startup budget. The result is Yeco Flash Mini-IT, a 118M-parameter Italian language model trained end to end for under €15, and published openly, limitations included.

This is not another commercial "wrapper", a graphical interface borrowing its brain via API from American servers. Yeco Flash Mini-IT is a proof of concept of a native proprietary LLM, built and trained entirely in-house.

The numbers of an economic and energy miracle

  • 118 million parameters: an ultra-light size conceived for fast logical reasoning in Italian
  • Trained for under €15: a reproducible pipeline that slashes academic and industrial development costs
  • Local CPU execution: no third-party cloud needed. Runs natively on very old consumer hardware at ~200 tokens per second

Measured on official benchmarks

On the official ITA-Bench suite, Yeco Flash Mini-IT answers BoolQ-IT at 62.5 vs 60.7 for Minerva-350M (figure from its paper), a model three times its size with ~70B pretraining tokens. Across the full suite the average still favors the larger model (36.1 vs 42.5), and we report that openly: the point of this release is a reproducible, honest baseline for sovereign Italian AI, not a leaderboard claim.

Sovereign stack, together with Yeco-PII

Paired with Yeco-PII, our 278M-parameter shield model that anonymizes sensitive data, beating OpenAI in six European languages with an average F1 of 0.966 vs 0.776 (historical v1.0 protocol), Yeco Flash Mini-IT is an infrastructure-grade alternative for enterprises that need digital sovereignty, designed with GDPR and EU AI Act requirements in mind (compliance depends on each deployment), with moderate inference costs.

Sovereign AI is no longer an abstract policy debate: it is an efficient, lightweight, independent engineering practice. Sometimes it fits in 118 million parameters.

The model is Apache 2.0 licensed and subject to rework as the series evolves. Read the full model card on its research page.

Related content