New app brings sovereign AI directly to our phones

Sovereign Danish AI for the general public has so far been in short supply. But Danes can now use a language model trained on Danish data that runs locally in an app on their phones. This comes as the Danish language model consortium Danish Foundation Models (DFM) today releases its “DFM Mimir” app.

The new Mimir v1.5 AI model, which powers the app, is trained on lawfully obtained Danish data and sends no data to foreign servers. It has taken the global lead among small language models with one billion parameters. This has generated considerable demand for the new app.

“Many people have asked whether they could run Mimir on their own devices easily and conveniently. Now they can. This means ordinary Danes can use Danish, sovereign AI completely free of charge, without sending their data abroad,” says Torben Blach, Programme Manager at Danish Foundation Models.

Smarter, but still very small

Mimir v1.5 has been trained by DFM researchers at the University of Southern Denmark. According to the evaluation site EuroEval, Mimir v1.5 now performs better in Danish than models 20 times its size. In English, it performs on a par with models four times its size.

This rapid development has been made possible in part by the extensive computing power available through the UCloud university collaboration. The model’s training data includes 125,000 documents from the Danish Parliament, recently made available by the Danish National Archives and the Danish Agency for Digital Government.

The result is a language model that can handle basic tasks for which many people would otherwise use foreign AI chatbots. These include summarising a text, performing calculations, creating a meal plan or writing a story.

“You can compare it to a person who has read more books. The model’s brain has not grown, but it knows more and makes better use of that knowledge. This shows that size is not what matters most when we develop language models. It is how we build them,” says Peter Schneider-Kamp, a professor at the University of Southern Denmark, who is part of DFM and is behind the Mimir series.

The limits have yet to be reached

The new model cannot yet search the web or take actions on the user’s behalf, and users will undoubtedly encounter other limitations. DFM therefore encourages users to report errors and shortcomings directly in the app so that the next version can be improved.

“We have not reached the limits of what a model of this size can do. That is what we are pursuing now, and we know it can still become significantly better,” says Peter Schneider-Kamp.

The “DFM Mimir” app is available to download free of charge for iPhone and for computers running Windows, macOS and Linux.