Swedish AI Models Worth Knowing About

Sweden has quietly built one of the more interesting AI ecosystems in Europe. Instead of chasing the biggest general-purpose model, researchers, public institutions and companies have focused on models that understand Swedish and the wider Nordic languages properly. That means better handling of compound words, local idioms, public-sector vocabulary and the kind of everyday text that global models often get slightly wrong. Here is a practical look at the main models and what each one does best.

The Models at a Glance

Model	Developer	Type	Languages	Best For
GPT-SW3	AI Sweden with RISE and WASP	Generative language model	Swedish, Norwegian, Danish, Icelandic, English	Text generation, chat, Nordic NLP research
KB-BERT	KBLab, National Library of Sweden	Encoder (BERT)	Swedish	Classification, named entity recognition, tagging
KB-Whisper	KBLab, National Library of Sweden	Speech recognition	Swedish	Transcription of meetings, media and archives
Swedish Sentence-BERT	KBLab, National Library of Sweden	Sentence embeddings	Swedish	Semantic search, clustering, similarity
Viking	Silo AI and TurkuNLP	Generative language model	Nordic languages, English, code	Multilingual Nordic applications
Llama-3 Swedish fine-tunes	AI Sweden and community contributors	Instruction-tuned LLM	Swedish, English	Local assistants and chatbots

GPT-SW3

GPT-SW3 is the model most people mean when they talk about a Swedish large language model. It was built by AI Sweden together with RISE and WASP, and it comes in a range of sizes, from small experimental versions of around 126 million parameters up to a 40 billion parameter model. It was trained on a large collection of Nordic text, and instruction-tuned variants exist for chat-style use.

Features

• Multiple sizes, so teams can pick a small model for cheap experiments or a large one for quality
• Trained on Swedish, Norwegian, Danish, Icelandic and English, with some code included
• Instruction-tuned versions for following prompts and holding conversations
• Designed with transparency in mind, including documentation of its training data

Strengths

• Natural, fluent Swedish output that avoids the translated feel of many global models
• Strong cross-Nordic ability, which helps with Norwegian and Danish content
• Useful as a base for fine-tuning on company or public-sector data
• A good fit for researchers who want to study how a regional model behaves

KB-BERT

Released by KBLab, the data lab at the National Library of Sweden, KB-BERT was one of the first serious Swedish-language transformer models. It was trained on a broad mix of Swedish text, including newspapers, government publications and online sources, which gives it a realistic grasp of how the language is written.

Features

• Standard BERT architecture, easy to load with common open-source libraries
• Cased vocabulary that respects Swedish capitalisation and compound words
• Freely available for fine-tuning on custom tasks

Strengths

• Reliable for named entity recognition, part-of-speech tagging and text classification
• Lightweight enough to run on modest hardware
• Proven in real workflows such as document sorting and information extraction
• A dependable baseline when a giant generative model would be overkill

KB-Whisper

Speech is where KBLab made a big leap. KB-Whisper takes OpenAI’s Whisper architecture and fine-tunes it on tens of thousands of hours of Swedish audio, much of it drawn from national archive material. The result is a transcription model that handles Swedish far better than the original multilingual version.

Features

• Available in several sizes, from small to large
• Trained on varied Swedish speech, including broadcasts, parliamentary recordings and conversational audio
• Produces punctuated, readable transcripts

Strengths

• Noticeably lower word error rates on Swedish compared with the base Whisper models
• Copes well with regional accents and formal as well as casual speech
• Practical for subtitling, meeting notes, podcasts and archive digitisation
• Open release makes it easy to run privately, which matters for sensitive recordings

Swedish Sentence-BERT

Also from KBLab, this model turns whole sentences into numerical vectors that capture meaning. It is not flashy, but it is the kind of tool that quietly powers a lot of useful software.

Features

• Produces fixed-size embeddings for Swedish sentences and short paragraphs
• Built on the KB-BERT foundation
• Simple to plug into vector databases and search pipelines

Strengths

• Finds documents by meaning rather than exact keywords
• Ideal for FAQ matching, duplicate detection and recommendation features
• Fast and cheap to run at scale
• A natural partner for retrieval-augmented generation setups that need Swedish context

Viking

Viking comes from Silo AI and the TurkuNLP group. It is a Nordic project rather than a purely Swedish one, but Swedish is a core language and the model is regularly used in Swedish settings. It was released in several sizes and trained on European supercomputing infrastructure.

Features

• Covers Swedish alongside Danish, Norwegian, Finnish, Icelandic and English
• Includes code in its training mix
• Released under a permissive open licence

Strengths

• Strong choice when one model has to serve several Nordic markets
• Open licensing makes commercial use straightforward
• Good balance between language coverage and manageable model size

Llama-3 Swedish Fine-Tunes

Not every Swedish model starts from scratch. AI Sweden and community developers have adapted Meta’s Llama 3 models to Swedish through further training and instruction tuning. This approach borrows the reasoning ability of a strong global model and adds local language skill on top.

Features

• Built on a modern, widely supported architecture
• Instruction-tuned for assistant-style conversations
• Compatible with popular local inference tools

Strengths

• Stronger general reasoning than many older Nordic models
• Runs on a single good GPU in its smaller versions
• Easy to deploy on private servers, which suits organisations with data-residency concerns
• Active community means frequent updates and new variants

Choosing the Right One

The right pick depends on the job rather than on size alone. For writing and chat, GPT-SW3 or a Llama-based fine-tune is the natural starting point. For sorting, tagging and extracting information from Swedish documents, KB-BERT is lean and dependable. If audio is involved, KB-Whisper is the clear leader for Swedish transcription. Search and recommendation features are best served by Sentence-BERT embeddings, and projects that span several Nordic countries can lean on Viking.

One practical tip: these models improve quickly, and new versions and benchmarks appear often. Before committing to a model for a production project, check the latest release notes and leaderboard results on the developers’ pages, and test the candidates on your own real Swedish text. A model that scores well on a benchmark can still stumble on your industry’s vocabulary, so a small trial with actual samples will tell you more than any ranking.