Overview
NbAiLab/borealis-12b-instruct-preview is a 12-billion parameter instruction-tuned model, released as a preview for early testing and feedback. Developed by NbAiLab, this experimental model is built upon the google/gemma-3-12b-it architecture and has been fine-tuned solely on textual instructions. It is part of a series of Borealis preview models, ranging from 270M to 27B parameters, with this 12B version offering a 32768 token context length.
Key Characteristics
- Base Model: Fine-tuned from
google/gemma-3-12b-it. - Training Data: Supervised fine-tuning (SFT) utilized the
NbAiLab/aurora-sft-2512 dataset. - Preview Status: This is an experimental, pre-release quality model, and users should expect unstable outputs and potential hallucinations.
- Safety Disclaimer: The model has not been safety-aligned and may produce harmful, biased, or insensitive content. It is not suitable for safety-critical applications without additional mitigations.
Intended Use Cases
- Norwegian-centric tasks: Optimized for assistant-style functions such as drafting, summarization, Q&A, and light reasoning in Norwegian.
- Writing Style Assessment: Useful for evaluating Norwegian writing style and quality.
- Early Evaluation: Designed for initial assessment of model behavior, language coverage (Norwegian, Bokmål, Nynorsk), and overall quality.
Available Formats
The model is provided in various formats, including original Transformers (safetensors), GGUF quantizations (q8_0, f16, bf16), and MLX formats (32-bit and 8-bit quantized) for Apple Silicon.