FRONTIER
GeneralKervalt-LM
General-purpose model for reading, reasoning, and answering across legal, financial, and scientific documents.
- Handles up to 128,000 tokens at once
- Hosted in the EU
- Zero public training on customer data
KERVALT / MODELS
Model research and model cards
We publish clear information about the models hosted on Kervalt. Every model card includes test results, data governance notes, and known limitations.
Open Model Cards
EU-Hosted Models
No Outside Training
Linked Sources
01 / HOSTED MODELS
Models you can use
Models you can run, adapt, and test without sending your data elsewhere.
CODE
TechnicalKervalt-Code
Built for code generation, architecture review, and working across Python, TypeScript, Go, and Java.
- Completes code in the middle of a file
- Understands long files
- Works with tools
MULTIMODAL
VisionKervalt-Vision
Reads documents and charts and turns them into structured, verifiable outputs.
- High-resolution document input
- Extracts tables in structured form
- Citations point to page regions
02 / EVALUATION
How we evaluate models
Our model research program is built on reproducible tests against the full Kervalt platform.
Tafari
Search benchmarks
What it does: Tests long-document recall, question answering, and citation fidelity across millions of pages.
Why it helps: Measures how well a model uses retrieved evidence rather than memory.
What you get: Buyers select models that perform on real document collections, not sanitized leaderboards.
Griot
Connection benchmarks
What it does: Tests connected reasoning, linking, and cascade prediction across documents.
Why it helps: Shows whether the model can follow connections across documents.
What you get: Risk and research teams identify models that avoid the blind spot.
Elimu
Record benchmarks
What it does: Tests joining records, date math, financial metric extraction, and schema adherence.
Why it helps: Forces numerical precision instead of guessing.
What you get: Compliance work gets reliable, checkable accuracy.
03 / SOVEREIGNTY
Your models stay under your control
Sending private questions to outside model services risks leaks and memorization. Kervalt keeps models inside your perimeter.
PICTURE
Data exposure
Sending private questions to outside model services risks leaks and memorization.
PROMISE
Self-hosted models
Models run on your European servers with no public fine-tuning and full audit logging.
PROVE
Published model cards
Every hosted model includes data provenance, test results, safety results, and known limitations.
PUSH
Request access
Talk to our research team about model access, custom tuning, and private benchmarks.
NEXT STEP
Evaluate models on your own data
Host models behind your perimeter and benchmark them against the full Kervalt platform.