NbAiLab/borealis2-26b-a4b-preview
NbAiLab/borealis2-26b-a4b-preview is a 26 billion parameter (4 billion active) instruction-tuned preview model developed by NbAiLab, based on google/gemma-4-26B-A4B-it. It features a 32768 token context length and is continuously pre-trained on the Aurora 2604 dataset, which includes 47 billion whitespace-separated words with a knowledge cutoff of January 1st, 2025. This model is specifically fine-tuned for Norwegian-centric assistant-style tasks, excelling in drafting, summarization, Q&A, and light reasoning, with a focus on Norwegian language coverage (Bokmål, Nynorsk, and English).
Loading preview...
Model Overview
NbAiLab/borealis2-26b-a4b-preview is a 26 billion parameter (4 billion active) instruction-tuned preview model from NbAiLab, designed for early testing and feedback. It is built upon the google/gemma-4-26B-A4B-it architecture and has undergone continuous pre-training on the extensive Aurora 2604 dataset, which comprises 47 billion whitespace-separated words with a knowledge cutoff of January 1st, 2025. The model was then supervised fine-tuned for two epochs using textual instructions from the NbAiLab/aurora-sft-2606 dataset, with SFT loss calculated exclusively on assistant responses.
Key Capabilities
- Norwegian Language Focus: Optimized for Norwegian-centric tasks, supporting Bokmål, Nynorsk, and English.
- Assistant-Style Tasks: Proficient in drafting, summarization, question answering, and light reasoning.
- Writing Assessment: Capable of assessing Norwegian writing style and quality.
- Large Context Window: Features a 32768 token context length.
Intended Use Cases
- Norwegian Assistant: Ideal for applications requiring an assistant-style model with strong Norwegian language capabilities.
- Content Generation: Suitable for drafting and summarization tasks in Norwegian.
- Language Evaluation: Useful for early evaluation of language coverage and quality, particularly for Norwegian.
Limitations
As a preview experiment, this model is not fully safety-aligned and may produce harmful or biased content. It is not recommended for safety-critical or high-stakes applications and should be used with caution, implementing additional safety mitigations if deployed. Outputs may be unstable and prone to hallucination, as this is an early SFT checkpoint rather than a final production model.