Lynx, a language model for the Nordic languages.Open weights, hosted in the EU.

Fine-tuned for Norwegian, Swedish, Danish and Icelandic and built on Qwen3-30B, with open weights you can run on your own servers.

Downloads on Hugging Face
1.2M
Tokens of context
262K
Base model
Qwen3-30B
Live across products
API + Chat

Where it started: NorskGPT

Highlights across the Nordic languages

EuroEval, March 2026. The base model, Qwen3-30B, covers 100+ languages.

Norwegian

NER (Nynorsk)71.0%
Common sense (Bokmål)69.3%
Summarization (Bokmål)63.4%

Danish

Civics knowledge79.3%
Reading comprehension72.2%
Common sense67.7%

Swedish

European values76.9%
Reading comprehension72.4%
Summarization65.9%

Icelandic

Summarization65.1%
NER63.6%
Reading comprehension58.6%

Built for use today, not just benchmarks

Nordic-first model behavior

Optimized for the kinds of Nordic-language interactions users actually have in products and daily work.

One model line across products

The same Lynx line powers both the Bineric API and Bineric Chat.

Built for practical deployment

Comes in 8-bit and 4-bit variants, sized for usable latency in real products.

Where Lynx is strongest

Selected highlights from our latest EuroEval runs, on a full 0 to 100 scale.

Danish Citizen Tests79.3%

Strong performance on Danish civics and knowledge tasks. MCC

VaLEU-sv — Swedish European Values76.9%

Strong alignment with Swedish European values benchmarks. MCC

MultiWikiQA (Swedish)72.4%

Solid Swedish reading comprehension on question answering. F1

MultiWikiQA (Danish)72.2%

Reliable Danish reading comprehension results. F1

NorNE Nynorsk71.0%

One of the strongest Norwegian named entity recognition scores in the current eval set. Micro F1

NorCommonSenseQA69.3%

Good common-sense reasoning on Norwegian tasks. MCC

How to read this

  • Selected highlight scores from our March 19, 2026 EuroEval (v16.11.0) runs across Norwegian, Swedish, Danish and Icelandic.
  • We show only the strongest areas, not a full leaderboard.
  • Tasks use different metrics (MCC, F1, BERTScore), so scores should be read as individual highlights rather than one combined ranking.

Key findings

  • Danish is the strongest language overall: 79.3% on Danish Citizen Tests.
  • Named entity recognition is strong where we measured it: 71.0% F1 on Nynorsk and 63.6% on Icelandic.
  • Reading comprehension leads in Danish and Swedish, both above 72% F1.
  • The 8-bit model outperforms 4-bit on every evaluated task: about 2% on average, up to 6.6% on Bokmål acceptability.

Ready to try Lynx?

Lynx is live across Bineric. Start in the chat, or build with it through the API.