Models

Viking — the Nordic open LLM from Kajaani

Silo AI's Viking is trained on 2 trillion tokens on the LUMI supercomputer in Kajaani. It speaks Finnish, Swedish, Norwegian, Danish and Icelandic natively — not via translation. Apache 2.0.

7B · 13B · 33BApache 2.0LUMI-trained

Viking is the first genuine Nordic frontier model. Silo AI released the family (7B, 13B, 33B) in 2024, continuing the Poro line and expanding to all Nordic languages plus programming languages. The OpenEuroLLM consortium is continuing the effort under EU funding.

Technical facts

Training data: 2 trillion tokens (FI, SV, NO, DA, IS, EN and code). Architecture: decoder-only transformer. Sizes: 7B, 13B, 33B. Licence: Apache 2.0 — commercial use allowed, fine-tuning allowed, the weights are yours. Infrastructure: EuroHPC LUMI in Kajaani.

What hardware runs Viking?

Viking 7B runs smoothly on a new MacBook Pro (M4 Pro/Max). Viking 13B wants a Mac Studio, Framework Desktop (Strix Halo) or equivalent. Viking 33B needs 128 GB unified memory (Mac Studio M-Ultra, DGX Spark). All three run offline.

Who should pick Viking?

You, if most of your work happens in Finnish or another Nordic language: leaders, researchers, lawyers, content creators and the public sector. Fine-tuning on your own domain (law, healthcare) yields the best results.

Frequently asked

How does Viking compare to Llama or Qwen?
On Finnish and Nordic languages, Viking is clearly better. On general reasoning, Llama 4 and Qwen 3.5 are ahead, but their Finnish is limited. A common setup is to run two models in parallel.
Can I use Viking commercially?
Yes. Apache 2.0 permits commercial use, redistribution and fine-tuning with no royalties.

Updated 2026-04-21

Want your own local AI assistant?

Tell us about your work and hardware — we'll map the right model, the right hardware tier and the right sync configuration.

Get in Touch