| <!doctype html> |
| <html lang="en"> |
| <head> |
| <meta charset="utf-8"> |
| <meta name="viewport" content="width=device-width, initial-scale=1"> |
| <title>BKS-Lab</title> |
| <style> |
| body { margin: 0; font: 15px/1.6 system-ui, -apple-system, "Segoe UI", sans-serif; color: #1A2433; background: #FFFFFF; } |
| main { max-width: 760px; padding: 8px 16px 24px; } |
| h1 { color: #0F2A4A; font-size: 24px; margin: 8px 0 12px; } |
| p { margin: 0 0 12px; } |
| ul { padding-left: 20px; margin: 0 0 12px; } |
| a { color: #2D5B8F; } |
| .de { color: #475569; border-top: 1px solid #E2E8F0; padding-top: 12px; } |
| @media (prefers-color-scheme: dark) { |
| body { color: #E7EAEE; background: #101419; } |
| h1 { color: #E7EAEE; } |
| a { color: #5690D2; } |
| .de { color: #A8B1BD; border-color: #333E4C; } |
| } |
| </style> |
| </head> |
| <body> |
| <main> |
| <h1>BKS-Lab</h1> |
| <p>We are a small team in Cologne that builds AI systems which run where our |
| customers' data is allowed to stay: on their own servers, or with a provider |
| under EU jurisdiction. Most of our work is making confidential company |
| knowledge searchable and usable for automation, with open models we can |
| host ourselves.</p> |
| <p>Before we recommend a local model, we measure it. Our test bench runs open |
| models on one RTX 4090 and compares them with hosted services on the same |
| tasks: German retrieval for RAG, German speech to text, invoice fields from |
| rendered page images, routing and extraction. Code, cases and raw rows are on |
| GitHub, the result tables are here.</p> |
| <ul> |
| <li>Results: <a href="https://huggingface.co/datasets/bks-lab/lab-bench-results" target="_blank">lab-bench-results</a> (dataset) and <a href="https://huggingface.co/spaces/bks-lab/lab-bench" target="_blank">lab-bench</a> (Space)</li> |
| <li>Invoice fields from page images: <a href="https://huggingface.co/datasets/bks-lab/xrechnung-invoice-fields" target="_blank">xrechnung-invoice-fields</a> (34 XRechnung test invoices with gold fields and our models' outputs) and its <a href="https://huggingface.co/spaces/bks-lab/xrechnung-invoice-fields-viewer" target="_blank">viewer</a></li> |
| <li>Models we measured, with our numbers: <a href="https://huggingface.co/collections/bks-lab/measured-on-one-rtx-4090-6aca0a9fa0d63e6bb9bc3b95" target="_blank">Measured on one RTX 4090</a> (collection)</li> |
| <li>Test bench on GitHub: <a href="https://github.com/bks-lab/lab-bench" target="_blank">bks-lab/lab-bench</a></li> |
| <li>Write-ups: <a href="https://bks-lab.com/en/labor/test-bench/" target="_blank">bks-lab.com/en/labor/test-bench</a></li> |
| <li>Our agent setup, open source: <a href="https://github.com/bks-lab/open-bridge" target="_blank">bks-lab/open-bridge</a></li> |
| <li>Why sovereign AI: <a href="https://bks-lab.com/en/souveraene-ki/" target="_blank">bks-lab.com/en/souveraene-ki</a></li> |
| </ul> |
| <p class="de"><b>Deutsch.</b> Wir bauen KI-Systeme, die dort laufen, wo die Daten bleiben |
| dürfen. Welches offene Modell eine Aufgabe trägt, messen wir vorher auf |
| eigener Hardware und legen die Messungen offen: |
| <a href="https://bks-lab.com/labor/pruefstand/" target="_blank">bks-lab.com/labor/pruefstand</a>.</p> |
| </main> |
| </body> |
| </html> |
|
|
