A 4-billion-parameter model that answers on Italian law - Software & AI
Software & AINovember 16, 2026

A 4-billion-parameter model that answers on Italian law

We published two models specialized on Italian law. Here are the numbers, how we measured them, and what they actually mean.

A 4-billion-parameter model that answers on Italian law - Software & AI | i3k

The problem with scores that have no context

Saying a model answers 89% of questions correctly means nothing on its own. If the starting model was already at 88%, the specialization did no work. If it was at 60%, it was a success. The number that counts is not the score: it is the gain.

So we publish both.

What we measured

We built an exam of 429 Italian-law questions, held out of training: the model never saw them while learning. That is a dull clarification and a decisive one, because a model evaluated on questions it has already met returns a meaningless score.

The results:

model            exam     base model     gain
legal-it-4b      85.1%    77.6%          +7.5
legal-it-8b      89.0%    80.9%          +8.1

After specialization, the four-billion-parameter model beats the unspecialized eight-billion one by four points. That was the result we cared about most: inside a narrow domain, specialization is worth more than size.

Why this changes the hardware arithmetic

A large generalist model needs datacenter GPUs. The two models we published weigh 2.5 GB and 5.2 GB as quantized GGUF Q4_K_M, and run on a workstation with a consumer GPU — in some cases on CPU alone.

For a company that moves the question from "what does AI infrastructure cost" to "which machine do I already have". It is the same reasoning that makes EuLLM Engine a single binary rather than a stack to orchestrate.

How they answer

They do not produce an answer from memory. They retrieve the relevant statutes and answer by citing them, so the reader can go back to the source and check. That is the difference between a system that asserts and one that shows its work: on legal material, only the second is useful.

It also makes them suited to sitting inside a RAG system rather than being used alone, which is exactly why we built them.

How they were trained

SFT followed by GRPO, on compute time allocated on the EuroHPC systems: Leonardo in Italy, LUMI in Finland and JUPITER in Germany. The base models are Qwen3-4B-Instruct and Ministral-3-8B, chosen because their licences place no branding constraints on derivatives.

What they are not

They are not a finished product and they do not replace a professional. They are two vertical models published to show the pipeline works and to let anyone check it: the weights are downloadable, the exam is repeatable, and the starting numbers are stated.

The next domain will not necessarily be legal. What is demonstrated is the method.

If you are weighing up work in this area, our private AI platform always starts from a real case, not a template.

FAQ

The models are on Hugging Face: legal-it-4b and legal-it-8b.

Interested?

Contact us to receive a personalized quote.

All articles

I3K Technologies Srl (single-member company) — i3k.eu