New Danish AI model challenges big tech rivals

Mimir, a Danish language model with 1 billion parameters, is on par with larger models developed by international tech giants.​

Mimir, named after the guardian of the well of wisdom in Norse mythology, was created by researchers at the University of Southern Denmark and the AI company Ordbogen A/S.

While it is limited to 1 billion parameters, data suggests its performance can be comparable to international models up to 5 times its size.

Parameters are variables that are learned by an AI model during training. The latest proprietary models, such as Claude Sonnet 5, are thought to have trillions of parameters.

Mimir was trained from scratch on just 8 GPUs in under 3 weeks. It used only legally and ethically obtained data, ensuring the training process was free of personal data and copyright infringements.

Peter Revsbech, CEO of Ordbogen A/S, says Mimir takes far less computational power to compete with the international tech giants.

“This makes digital sovereignty for Denmark and the EU more than a dream. It's a plan. And we are in the process of executing it,” Revsbech wrote in a LinkedIn post.

Mimir can be downloaded here and run on Google Colab, Kaggle, and some local apps.

How does Mimir compare to its rivals?

Researchers evaluated Mimir against a broad range of benchmarks in English, Math & Code, and Danish. The findings are presented in the preprint paper published in arXiv.

Mimir was compared with other 1B-parameter models, such as HRM-Text by Singapore-based Sapient Intelligence, Gemma 3 by Google DeepMind, and OLMo2 by the Allen Institute of AI from Seattle.

In addition, Mimir was compared with larger models in the 2-3B and 4-5B ranges, including Gemma 4 E2B, which has 5B total parameters.

On average, Mimir outperformed all 1B-model competitors across all English benchmarks and was only 0.3 points behind China’s Qwen 3.5, which has 4B parameters.

The Danish model also showed strong performance on Math and Code benchmarks, outperforming all other international models of the same size.

Among larger models, it ranked second overall on the GSM8K benchmark, which tests a model’s ability to perform multi-step mathematical reasoning.

On the HumanEval benchmark, which assesses the model's code-generation capabilities, Mimir outperformed China’s Qwen3.5-2B.

In addition, Mimir demonstrated the strongest overall performance on Danish benchmarks, even beating models up to 8-9 times larger, such as Munin-Apertus 8B, developed under the Danish Foundation Models initiative.

Can Europe build its own AI models?

Creators say that Mimir is proof that Denmark and the EU can develop and operate competitive AI models themselves, which could enable public institutions and businesses to operate without sensitive data leaving the continent.

Mimir comes as Europeans are actively working to reduce their technological dependence on foreign countries like the US and China, with data privacy at the forefront of the sovereignty push.

However, Mimir is far from being the first European AI model.

French company Mistral AI builds its own proprietary and open-weight large language models (LLMs) – much bigger than Mimir – that are already competing with American and Chinese counterparts.

The German research consortium released Soofi S, an open-source German-English AI model, in July 2026. The 30B parameter model outperformed some of its European and US open-source rivals in English, German, coding, maths, and regional knowledge.