| <!doctype html> |
| <html lang="en"> |
| <head> |
| <meta charset="utf-8"> |
| <meta name="viewport" content="width=device-width, initial-scale=1"> |
| <title>EuLLM</title> |
| <style> |
| body { font-family: system-ui, -apple-system, "Segoe UI", sans-serif; max-width: 860px; margin: 0 auto; padding: 16px; line-height: 1.5; color: #1f2937; background: #fff; } |
| h1 { font-size: 1.4rem; margin: 0 0 .2rem; } |
| h2 { font-size: 1.1rem; margin: 1.4rem 0 .4rem; } |
| p { margin: .4rem 0; } |
| table { border-collapse: collapse; width: 100%; font-size: .92rem; margin: .4rem 0; } |
| th, td { border: 1px solid #e5e7eb; padding: 6px 8px; text-align: left; vertical-align: top; } |
| th { background: #f9fafb; } |
| ul { margin: .3rem 0; padding-left: 1.2rem; } |
| a { color: #2563eb; } |
| .tag { font-weight: 600; } |
| @media (prefers-color-scheme: dark) { |
| body { color: #e5e7eb; background: #0b0f19; } |
| th { background: #111827; } th, td { border-color: #374151; } a { color: #60a5fa; } |
| } |
| </style> |
| </head> |
| <body> |
| <h1>EuLLM — sovereign LLMs for Europe</h1> |
| <p class="tag">Modelli linguistici aperti, verticali e locali, costruiti in Europa.</p> |
| <p>EuLLM builds open-weight language models specialised for one domain and one language, small enough to run on your own hardware, and the engine to run them. No data leaves your machine.</p> |
|
|
| <h2>Models</h2> |
| <table> |
| <tr><th>Model</th><th>What it does</th><th>Size (Q4_K_M)</th><th>Held-out exam</th></tr> |
| <tr><td><a href="https://huggingface.co/EuLLM/legal-it-8b" target="_blank">legal-it-8b</a></td><td>Answers questions on Italian law from the statutes retrieved for it, citing them</td><td>5.2 GB</td><td>89.0% (base model: 80.9%)</td></tr> |
| <tr><td><a href="https://huggingface.co/EuLLM/legal-it-4b" target="_blank">legal-it-4b</a></td><td>The same, for smaller machines</td><td>2.5 GB</td><td>85.1% (base model: 77.6%)</td></tr> |
| </table> |
| <ul> |
| <li><b>How they are used.</b> Both are built for retrieval-augmented, open-book use. The recommended retrieval (BM25, Qwen3-Embedding-0.6B and Qwen3-Reranker-0.6B, about 1.2 GB in all) is documented in the model cards.</li> |
| <li><b>How they were measured.</b> On 429 held-out questions never seen in training, with paired significance tests and with a check that needs no judge model.</li> |
| <li><b>Not legal advice:</b> they are tools for consulting the legal texts.</li> |
| </ul> |
|
|
| <h2>Engine</h2> |
| <p><a href="https://github.com/eullm/eullm" target="_blank">EuLLM Engine</a> is a single Rust binary for running models like these locally:</p> |
| <ul> |
| <li>the same API as Ollama and OpenAI, so existing clients work unchanged;</li> |
| <li>continuous batching;</li> |
| <li>no telemetry;</li> |
| <li>an audit trail designed for the EU AI Act.</li> |
| </ul> |
|
|
| <h2>Built with European compute</h2> |
| <p>The models are trained on the Leonardo supercomputer (CINECA, Italy) through the EuroHPC AI Factory allocation EHPC-AIF-2026PG01-1147. We thank the EuroHPC Joint Undertaking for access.</p> |
|
|
| <p><a href="https://www.eullm.eu" target="_blank">eullm.eu</a> · <a href="https://github.com/eullm" target="_blank">GitHub</a></p> |
| </body> |
| </html> |
