A small Danish language model aims to provide greater control over AI.

When staff and students use artificial intelligence, it is often through services provided by large foreign companies. But at SDU, researchers are also working on developing the technology itself.

One of the results is Mimir, a small language model developed as part of the Danish Foundation Models. It has been trained from scratch with a particular focus on the Danish language and knowledge of Danish contexts and can be downloaded and run on a local computer or an organisation’s own server.

Professor Peter Schneider-Kamp from the Department of Mathematics and Computer Science discusses here the model’s capabilities and limitations, and why Danish language models are an important area of research.

1. What is Mimir, and what makes this model special?

A language model is the technology behind many of the chatbots that can answer questions and generate text.

– We have succeeded in creating a small model that performs better than the leading models of the same size from Google and Alibaba. In fact, it is on a par with models that are far larger. So, that is a major achievement.

According to Schneider-Kamp, Mimir is particularly strong at Danish grammar. In tests of grammatical correctness and the correction of grammatical errors, it performs best among all the models the research team has compared it with, regardless of size.

A smaller model may require fewer computing resources, thereby making it possible to run it on a local computer, for example to correct language or summarise a conversation.

– Development has been characterised by ever-larger models with high resource consumption. Now we are seeing that smaller models can handle some of the same tasks. I was personally surprised at how well Mimir performs these tasks.

2. Why develop a Danish language model?

– From the outset, we were concerned with data sovereignty. This is about having control over where our data is processed and what data we use to train the models. At the same time, we wanted to ensure that the Danish language, culture and values were represented.

– The most important thing is that we have trained the model from scratch and exclusively with data that we are permitted to use for this purpose. We have personally overseen the training and selection of data, with a focus on copyright and the protection of personal data.

For Schneider-Kamp, the work is also about access. In his view, Denmark and Europe cannot take it for granted that they will always have access to the largest and latest foreign models. That is why it is important to develop alternatives that others can explore, adapt and build upon.

3. What can Mimir be used for right now?

Schneider-Kamp points to correcting spelling and grammar errors and summarising conversations and meeting minutes as specific applications. These are tasks that are relevant at universities, in businesses and within public authorities.

Mimir can also answer questions, write texts, perform calculations and programme. However, it cannot compete with the largest models in all areas.

– One should not expect Mimir to be able to compete broadly with the latest models behind ChatGPT and Claude right from the start. But in some areas, it is already better than the original ChatGPT, including for performing calculations and programming.

Mimir does not have the same breadth of general knowledge as the large models, but it can work with information provided by the user. It can also be integrated into solutions where a search function retrieves up-to-date information from the internet, which the model can then use in its responses.

Mimir can also be run locally. If both the model and the programme in which it is used run on the computer and do not send conversation data to external services, the information remains on the computer. The model is freely available, but there are, of course, ongoing costs for equipment, electricity and operation.

4. How might AI change the way work is done at SDU?

Schneider-Kamp expects AI to become more integrated into the systems that staff and students already work in.

– At present, much of its use takes place via chatbots. You copy material into it, get a reply and then copy it on to another programme. In five years’ time, I think AI will be less visible as a standalone tool.

He envisages, for example, a biologist asking a system to investigate a correlation in research data and having the result displayed as figures or text. This requires both suitable models and an infrastructure where the data can be processed securely.

This is where UCloud comes in. Mimir has been trained using computing resources from UCloud, and UCloud AI provides access to language models on Danish infrastructure. The platform can be used to work with sensitive data, provided the application falls within the approved framework.

– We need good models that are freely available and that we can run locally, so that we are not dependent on sharing our data.

5. What is the next step for Mimir?

– We expect the next version to be able to handle more European languages. Looking ahead to 2027, we are also working on a larger version, which we expect will be able to compete with medium-sized models.

Companies such as Ordbogen A/S, which have contributed to the development of Mimir, are simultaneously working on solutions that build on these models.

– At Danish Foundation Models, our main task is to develop open, well-documented language models that are trained on data we are permitted to use, and which others can build upon. We are laying the foundations that companies and other stakeholders can use to develop concrete solutions.

Our ambition is to develop the best European language model and demonstrate that Europe can play a part in this development.

– We have limited resources, and this forces us to find alternative approaches. With Mimir, we have shown that we can develop a competitive model without a billion-budget.