NbAiLab/borealis-12b

Borealis 12B: Norwegian-Centric Instruction-Tuned Model

NbAiLab/borealis-12b is a 12-billion parameter instruction-tuned model developed by the National Library of Norway (NbAiLab). It is part of the Borealis family, built upon the Gemma 3 architecture, specifically fine-tuned from google/gemma-3-12b-it.

Key Differentiators & Capabilities

• Norwegian-Centric Fine-tuning: This model is uniquely instruction-tuned for Norwegian language tasks, including assistant-style interactions, writing, summarization, and question answering.
• Unique Training Data: It incorporates a subset of data from an agreement between Norwegian rights-holder organizations and the government, including copyrighted press material. This data specifically teaches the model to generate news article titles and ingress texts, a feature not found in the -open Borealis variants.
• Safety Alignment: Utilizes a technique called prompt baking to distill system prompt behaviors directly into the model weights, balancing quality, usefulness, and safer behavior.
• Multilingual Support: While primarily Norwegian-centric, it also supports English use cases, with performance characterized for Norwegian, Bokmål, and Nynorsk.

Intended Use Cases

• Norwegian Assistant Tasks: Ideal for drafting, summarization, and Q&A in Norwegian.
• Writing Assessment: Useful for evaluating and improving Norwegian writing style and quality.
• Language Coverage Evaluation: Suitable for assessing behavior and language coverage across Norwegian, Bokmål, and Nynorsk.

Limitations

• May produce incorrect or hallucinated information.
• Safety alignment reduces, but does not eliminate, harmful or inappropriate outputs.
• Performance outside Norwegian and English has not been fully characterized.