Hopp til hovedinnhold
tenki
All models
The deep oneExperimental

Hugin

Reasons its way to an answer and takes longer doing it. Named for the raven Odin sent out to think.

· Specifications

Model
Our own Norwegian fine-tune, trained by Tenki
Size
4 billion parameters, 4-bit quantised MLX build
Context window
262,144 tokens native, practical limit lower on 16 GB
Runs on
Mac mini M4 running mlx_lm.server, located in Norway
Interface
OpenAI-compatible chat completions
Access
Free and anonymous at /chat, or with a token via the API
Conversation storage
None. Conversations are not stored.
Profile
Longer reasoning, slower answers
01

What Hugin is

Hugin is a public endpoint we operate ourselves. Underneath it is our own Norwegian fine-tune, trained by us. It runs on a Mac mini located in Norway, on hardware we own, and answers without the request passing through a third party.

Hugin and Munin are the same fine-tuned model. The difference between them is how we run them: how much room the model gets to reason before answering, and how the answer is retrieved. That is a configuration difference, not two sets of weights.

02

Intended use

Hugin exists to be tried. It is free, requires no sign-in, and is there so people can get a feel for what a local model actually manages before deciding anything.

  • · Exploring what local AI is, and where its limits sit
  • · Questions about AI, privacy, the AI Act, and how we work
  • · Drafts, rewrites, and explanations where you read them over yourself
03

Out of scope

Hugin is experimental and should not stand alone in anything that has consequences for a person. We mean that literally: do not use it as the sole basis for a decision.

  • · Legal, medical, or financial advice
  • · Processing personal data or client data
  • · Decisions about hiring, credit, health, or benefits
  • · Anything requiring a verifiable answer without a human reading it
04

Limitations

Hugin is a small model. Four billion parameters, quantised to four bits to fit the hardware we run it on. That is a deliberate trade: we prioritise the model sitting in Norway on hardware we own over it being as large as possible. The fine-tune recovers some ground on Norwegian, but it does not undo the size. The consequence is that it loses to the large cloud models on tasks needing broad factual knowledge, long chains of reasoning, or precise code.

  • · It makes things up. Verify anything that will be used for something that matters.
  • · Its knowledge is general and has a cutoff. It does not know your internal documents.
  • · Norwegian bokmål is weaker than English, and nynorsk weaker still.
  • · Arithmetic, dates, and statutory references are among the weakest areas.
  • · It remembers nothing between conversations, because we do not store them.
05

Data and privacy

Conversations at /chat are not stored. There is no account, no history, and no conversation log to retrieve later, including for us.

We store aggregate usage (which model, how many requests) and a thumbs up or down if you give one, because that feedback is used to pick better examples. Requests are rate limited per IP address and pass through Cloudflare Turnstile to keep automated abuse down.

06

Evaluation

We run our own benchmark, TenkiBench, with tasks drawn from Norwegian SMB work: invoices, contracts, VAT, statutory references, business registry lookups, HR and payroll, customer service, and bokmål against nynorsk. The tasks, the weighting, and the evaluation methods are public, and model vendors do not pay to be measured.

We run it on ourselves too. Where Hugin loses to a larger cloud model, it says so in the table.