Viking
Open source language model for Nordic languages and code
Viking is a family of open source large language models developed by the TurkuNLP group, SiloGen and High Performance Language Technologies. Released from 2023, it focuses on Finnish, English and Scandinavian languages to support low resource linguistic regions.
- Trained on 2 trillion tokens of Nordic languages and code
- Built on the LUMI supercomputer using AMD MI250X GPUs
- Released under the permissive Apache 2.0 open source licence
- Supports Finnish, Swedish, Danish, Norwegian and Icelandic
- Uses a custom 128K token vocabulary for better language coverage
What it is
Viking is a set of generative pretrained transformer models designed for Nordic languages and programming code. The largest version, Viking 33B, contains 33 billion parameters. It was built by the TurkuNLP group at the University of Turku, SiloGen from Silo AI, and High Performance Language Technologies. The project aims to provide open source tools for languages with fewer digital resources, such as Finnish, Icelandic and the Scandinavian tongues. It serves as a base model that requires further tuning for specific applications.
How it came about
Training for the Viking family began in September 2023. The project represents the second set of models released by the LumiOpen collaboration. Checkpoints were made available progressively as training advanced, with releases every 100 billion tokens processed. The initiative involves partners from Finland and the wider Nordic region, leveraging local high performance computing infrastructure to develop these linguistic tools.
Inside the model
The model uses a decoder-only transformer architecture similar to LLaMA, incorporating rotary positional embeddings and flash attention. It was trained on the LUMI supercomputer using 1024 AMD MI250X graphics processing units. The training setup employed a custom fork of the Megatron-Deepspeed framework with a three-dimensional parallelism strategy. A custom tokenizer with 128,000 entries was created specifically for the mixed dataset of Nordic languages, English and code.
What it can do
Viking is fluent in Finnish, English, Swedish, Danish, Norwegian and Icelandic. It can perform basic translation between these languages and understands programming code. The developers state it has no meaningful proficiency in languages outside this specific set. As a base model, it generates text but needs fine-tuning to align with specific user needs or ethical standards. Full evaluation results were pending publication at the time of the release notes.
How to use it
The model weights are fully open source under the Apache 2.0 licence. Users can download the models from Hugging Face. The repository includes various training checkpoints, allowing developers to select versions based on the amount of data processed during training. The main branch always points to the latest available checkpoint, which reached 1.3 trillion tokens at the time of the documented release.
Safety and testing
The developers warn that Viking is a partially trained product and outputs should be treated with care. Like other systems trained on web data, it may reflect biases, inaccuracies or controversial views present in its training corpus. Users are advised to exercise discretion and conduct additional evaluations to ensure responses meet their ethical requirements. No specific red-teaming results were detailed in the provided documentation.
Why it matters
Viking contributes to the ecosystem of open source models for non-English languages. By focusing on Nordic tongues, it addresses the gap in high quality language technology for these regions. The use of the LUMI supercomputer demonstrates the capability of European research infrastructure to support large scale artificial intelligence development without relying solely on commercial cloud providers.
Questions people ask about Viking
What languages does Viking support?
Viking is optimized for Finnish, English, Swedish, Danish, Norwegian, Icelandic and programming code. The developers state it has no meaningful proficiency in other languages.
Is Viking free to use?
Yes, the model weights are fully open source and released under the Apache 2.0 licence, allowing for broad usage and modification.
Can I use Viking directly for chat?
No, Viking is a base model. It requires further fine-tuning and alignment before it is suitable for most interactive use cases like chatting.
Good reading and sources
There is no English Wikipedia article on Viking yet, so the facts on this page come from the sources listed above: huggingface.co/LumiOpen/Viking-33B. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.