Hopp til hovedinnhold
tenki
NO

· Agenda

What we work on

Four questions we spend our time on. Each has something running behind it, not just an opinion. Where the work is unfinished, it says so.

01

Local inference on hardware people actually own

The questionHow much use can you get from a model small enough to sit on your own premises?

What is running

  • · Hugin and Munin, two public endpoints running on a Mac mini M4 in Norway, free and without sign-in.
  • · A retrieval layer around the model that does more for answer quality than the parameter count does: topic routing with local embeddings, and examples picked from what users actually gave a thumbs up.

What is still open

A 4-bit quantised model at four billion parameters loses to the cloud models on most things. The open question is where it is nonetheless good enough, and how much of the gap can be closed with retrieval and context rather than larger weights.

The model cardsTry them
02

Measurement in Norwegian, not translated measurement

The questionDoes a language model work on Norwegian office work, or only on English benchmarks?

What is running

  • · TenkiBench: 116 tasks drawn from Norwegian SMB work, spread across invoices, contracts, VAT, statutory references, business registry lookups, HR and payroll, customer service, and bokmål against nynorsk.
  • · The tasks, the weighting, the evaluation methods, and a hold-out set with published hashes are all open under CC BY 4.0, with ready-made citations.
  • · Model vendors do not pay to be measured, and there are no ads on the result tables.

What is still open

External expert validation of the tasks is being built, not finished. We are recruiting lawyers, auditors, accountants, and linguists to sign off that the tasks genuinely reflect their work, and the tasks do not yet carry such a signature.

TenkiBenchMethodologyBecome a validator
03

Regulation made searchable

The questionCan a small company find out what it is allowed to do without paying for an hour of legal advice first?

What is running

  • · A search over decisions from the Norwegian Data Protection Authority, appeal cases from the Privacy Appeals Board, and guidance from Digdir, where you write in plain Norwegian and get hits on meaning rather than keywords.
  • · A risk classifier that places an AI system in the AI Act's four tiers, and a tool that drafts a DPIA following the Data Protection Authority's own template.
  • · All of it free and without sign-up.

What is still open

The tools produce drafts and surface sources. They do not replace a lawyer, and we say so on each of them.

Tenki ComplianceThe AI Act for SMBs
04

Agentic systems put into operation

The questionWhat happens when language models get tools, memory, and permission to act over time?

What is running

  • · We run our own internal system of digital colleagues with their own memory, their own tool boundaries, and an approval queue where a human has to say yes before anything that changes data is carried out.
  • · Tool access goes through our own MCP servers, with a separate token per colleague rather than one shared key.

What is still open

This runs internally with us, not with clients. We think that is the right order: we run it on our own operations long enough to know where it fails before proposing it to anyone else.

How we work

The agenda is deliberately narrow. We are a small company, and a small company working on everything is really working on nothing.