Choosing a Nordic AI model? Here’s what to check first

Finding an AI model that supports Swedish or Finnish is a useful start. Working out whether it fits your project takes a little more digging.

Can you download it? Can you use it commercially? Will it work as a chatbot, or does it need more development?

Viking and GPT-SW3 show why these questions matter. Both are established Nordic language model projects, but their documentation tells different stories about access and practical use.

Start with the current access information

An announcement can remain online long after the situation it describes has changed.

AI Sweden’s November 2023 announcement described an open release of GPT-SW3 for commercial and non-commercial use. Its current project page, however, says the models are no longer available for open download. It describes a research-access application process for people affiliated with European universities or research institutes. Applications are reviewed individually. [1][2]

The current page does not explain when that arrangement changed, so this should not be read as a new announcement.

Before building plans around a model, visit its current project page and check the exact download and access conditions. An old article may accurately describe the launch while giving you the wrong impression about availability today.

Check what you’re actually downloading

Viking 33B’s official model card describes it as a base model that needs further fine-tuning for most uses. The card lists an Apache 2.0 licence and describes training involving Finnish, English, Swedish, Danish, Norwegian, Icelandic and code. [3]

That distinction matters if you expect a finished conversational assistant.

A base model, an instruction-tuned version and a community adaptation may share a name, but they are different things to evaluate. Check the exact repository, version and accompanying documentation before you download.

If you choose an adaptation made by another developer, read its model card as well as the original one.

Try the language and task you actually need

A list of supported languages cannot tell you whether a model will handle your work well enough.

Imagine a Finnish retailer testing an assistant for customer enquiries. A reply might sound natural while misunderstanding a return request or inventing a shop policy.

A useful first test does not have to be complicated. Start with around 20 realistic examples using public, synthetic or appropriately authorised material. Write down what a good answer should contain before testing the model.

Include spelling mistakes, local names, mixed-language messages and questions that cannot be answered from the information provided.

Ask a fluent speaker to review the results. Assess whether the answers are correct as well as whether they sound natural. Keep examples of failures so you can compare versions and see whether changes help.

This is our suggested starting method, not an independent benchmark or a claim about either model’s performance.

Think beyond the download

Downloading weights is only one part of running an AI system.

Before committing time or money, work out how you will host it, maintain it and review its output. Check the permissions attached to your chosen checkpoint and any additional terms for the service you plan to use.

European development or Nordic training infrastructure alone does not tell you where a separate hosting provider will process your data. Check the actual deployment arrangements.

Keep a short record of your decision

For each model on your shortlist, note the exact checkpoint, access conditions, licence, review date and results from your own tests.

You should be able to explain why you selected it and what remains uncertain. That record will be more useful than choosing on the strength of a model name or a single impressive answer.

Frontier AI’s comparison tool can help you organise documented facts: https://frontierai.fi/compare/

It does not provide an independently measured performance ranking. Use it to build your shortlist, then follow the original documentation and test the exact version you intend to use.

Sources reviewed on 21 October 2026. Access arrangements and documentation may change. Frontier AI has not independently benchmarked these models.

References: [1] AI Sweden’s current GPT-SW3 project page; [2] AI Sweden’s November 2023 open-release announcement; [3] LumiOpen’s official Viking-33B model card. Links are provided in the sources section.