borealis2-26b-a4b-preview
Public
1 Download
Capabilities
Minimum system memory
Tags
1 Download
Capabilities
1 Download
Capabilities
Minimum system memory
Tags
1 Download
Capabilities
Minimum system memory
Tags
Last updated
Updated 3 days agobyREADME
Norwegian-centric instruction-tuned model from the AI Lab at the National Library of Norway
(Nasjonalbiblioteket). Continued pre-training + SFT on top of google/gemma-4-26B-A4B-it.
26B total parameters, ~4B active per token.
Quantizations by BobTheShoplifter — format conversions only, nothing retrained.
| Variant | Size | Notes |
|---|---|---|
| GGUF Q4_K_M | 16.8 GB | Norwegian-calibrated imatrix. Runs anywhere llama.cpp runs. |
| GGUF Q6_K | 22.6 GB | Norwegian-calibrated imatrix. |
| GGUF Q8_0 | 26.9 GB | Effectively lossless. |
| MLX 6-bit | 20.5 GB | Apple Silicon, fastest of the set. |
| MLX 8-bit | 26.8 GB | Apple Silicon, effectively lossless. |
Images / OCR: supported by the GGUF variants only, via the mmproj file in that repo. The MLX
builds are text-only — mlx-vlm has no gemma4 support yet, so there is no MLX vision path to ship.
Defaults here ship the model's own sampling settings (temp 1.0 / top-k 64 / top-p 0.95).
Reasoning: use a GGUF variant. LM Studio's llama.cpp runtime supports Gemma 4's reasoning channel and folds it into a collapsible block automatically. Its runtime does not — MLX variants answer directly and cannot be made to reason (measured: GGUF Q6_K 426 reasoning tokens, MLX 8-bit 0, same prompt). This affects every gemma4 MLX model in LM Studio, not just this one. MLX is still faster (94–110 vs 75–109 tok/s) and better at Norwegian OCR.
Custom Fields
Special features defined by the model author
Enable Thinking
: boolean
(default=true)
Let the model write a reasoning block before answering. Borealis 2 tends to reason in English even when answering in Norwegian, and often reasons anyway on hard prompts. Off matches the upstream chat template default.
Parameters
Custom configuration options included with this model
Sources
The underlying model files this model uses
Minimum system memory
Tags
Last updated
Updated 3 days agobyREADME
Norwegian-centric instruction-tuned model from the AI Lab at the National Library of Norway
(Nasjonalbiblioteket). Continued pre-training + SFT on top of google/gemma-4-26B-A4B-it.
26B total parameters, ~4B active per token.
Quantizations by BobTheShoplifter — format conversions only, nothing retrained.
| Variant | Size | Notes |
|---|---|---|
| GGUF Q4_K_M | 16.8 GB | Norwegian-calibrated imatrix. Runs anywhere llama.cpp runs. |
| GGUF Q6_K | 22.6 GB | Norwegian-calibrated imatrix. |
| GGUF Q8_0 | 26.9 GB | Effectively lossless. |
| MLX 6-bit | 20.5 GB | Apple Silicon, fastest of the set. |
| MLX 8-bit | 26.8 GB | Apple Silicon, effectively lossless. |
Images / OCR: supported by the GGUF variants only, via the mmproj file in that repo. The MLX
builds are text-only — mlx-vlm has no gemma4 support yet, so there is no MLX vision path to ship.
Defaults here ship the model's own sampling settings (temp 1.0 / top-k 64 / top-p 0.95).
Reasoning: use a GGUF variant. LM Studio's llama.cpp runtime supports Gemma 4's reasoning channel and folds it into a collapsible block automatically. Its runtime does not — MLX variants answer directly and cannot be made to reason (measured: GGUF Q6_K 426 reasoning tokens, MLX 8-bit 0, same prompt). This affects every gemma4 MLX model in LM Studio, not just this one. MLX is still faster (94–110 vs 75–109 tok/s) and better at Norwegian OCR.
Custom Fields
Special features defined by the model author
Enable Thinking
: boolean
(default=true)
Let the model write a reasoning block before answering. Borealis 2 tends to reason in English even when answering in Norwegian, and often reasons anyway on hard prompts. Off matches the upstream chat template default.
Parameters
Custom configuration options included with this model
Sources
The underlying model files this model uses
This is a preview experiment, not a production model. It has not been fully safety-aligned and may follow harmful instructions. Outputs may be unstable and may hallucinate. Do not use it for safety-critical or high-stakes applications.
NB-License 1.0 — Apache 2.0 adapted with additional use-based restrictions. You must not use the model to intentionally recreate its training data, nor use it or its output to power services whose primary purpose is giving access to licensed press publications in that training data.
Model by NbAiLab — https://ai.nb.no · https://huggingface.co/NbAiLab/borealis2-26b-a4b-preview
Based on
This is a preview experiment, not a production model. It has not been fully safety-aligned and may follow harmful instructions. Outputs may be unstable and may hallucinate. Do not use it for safety-critical or high-stakes applications.
NB-License 1.0 — Apache 2.0 adapted with additional use-based restrictions. You must not use the model to intentionally recreate its training data, nor use it or its output to power services whose primary purpose is giving access to licensed press publications in that training data.
Model by NbAiLab — https://ai.nb.no · https://huggingface.co/NbAiLab/borealis2-26b-a4b-preview
Based on