Skip to content
Skjld Labs

Free Norwegian language models are here: how your business can use Borealis

Borealis is a family of open language models from the National Library of Norway, built for Norwegian Bokmål and Nynorsk. They are free to download, and you can run them yourself. Here is what the models are, which size fits which job, and how to get started.

9 min read Laget med AI-hjelp
An aurora-like band of text above mountains and a fjord, with a laptop in the foreground

In short

Borealis is a family of open language models from the National Library of Norway's AI Lab. They are built for Norwegian Bokmål, Nynorsk and English, and they are free to download from Hugging Face. You can run them on your own machine or your own server, so the text you send in stays with you. The models come in five sizes, from 270 million to 27 billion parameters.

What Borealis is

Borealis arrived in two rounds in 2026.

The first family was presented on 26 May. It builds on Google's open Gemma 3 and is adapted to Norwegian for tasks such as writing, summarising and question answering. It comes in five sizes: 270M, 1B, 4B, 12B and 27B.

On 10 August, the National Library launched the first models where text from Norwegian newspapers is used in training. The use is governed by a collective agreement between the National Library and Kopinor on behalf of Norwegian newspapers, and the government has allocated NOK 45 million a year to it. Karianne Tung, Minister of Digitalisation and Public Governance, said at the launch that hospitals, the police and Norwegian businesses are getting a new and important tool.

The next generation is called Borealis 2 and builds on Gemma 4. It is on Hugging Face as a 31-billion-parameter preview with a release date of 28 September 2026. The model card itself describes it as an early SFT checkpoint, not a final production model.

The name comes from Latin and means "northern", as in aurora borealis, the northern lights.

Which size should you choose?

Larger models write better and know more, but they ask more of the machine. The figure below shows the file size of the most used compressed version (Q4_K_M) from the National Library on Hugging Face, read on 19 November 2026, and a rule of thumb for hardware.

Borealis: five sizesFile size for the compressed Q4_K_M build, and a rule of thumb for the machine.270M253 MBOrdinary laptopTrials and very simple tasks1B806 MBOrdinary laptopSmall tasks with few resources4B2.49 GBRecent PC or MacA good choice when cost matters12B7.3 GBGraphics card with 12 GBGeneral work in Norwegian27B16.5 GB24 GB graphics card, or a serverBest quality in the familySource: NbAiLab on Hugging Face, read 19.11.2026.
The model has to fit in memory, and it needs some extra room for the text it is working on. Expect to need somewhat more than the file size.

The National Library itself highlights 12B and 27B as the strongest for general use. They also write that 4B is surprisingly capable for its size, and a good practical choice when running costs matter.

What the 27B needs in hardware

The largest model is the one that most often stalls people, so here are the numbers. It needs room for the weights themselves, plus memory for the context it is working on (KV cache) and for the runtime.

What a 27B takes upThe weights are only one part. Context and runtime need room too.Model weightsNVFP4 approx. 15.2 GBKV cacheand activationsRuntime16 GB cardjust fits the weights24 GB cardgood margin010 GB20 GB28 GBRecommended total: 20–24 GB VRAMThe Q4_K_M file is 16.5 GB. The NVFP4 weights are approx. 15.2 GB. Both point to a 24 GB card.

If you run the model in NVFP4, the format uses 4 bits per parameter (0.5 bytes), plus a small overhead for block scaling where E4M3 scales per 16 values add around 0.5 bits per parameter. That gives 27 billion × 0.5625 bytes ≈ 15.2 GB for the weights alone.

A 16 GB card, such as the RTX 4060 Ti 16GB, RTX 5060 Ti 16GB or RTX 4080, just fits the weights. But that leaves little room for KV cache and activations, and you quickly risk Out of Memory on longer conversations or larger batch sizes. With 24 GB (RTX 3090 or 4090) or 32 GB (RTX 5090, Radeon AI PRO R9700) there is ample room for a large context window of 32k–64k tokens or more, especially with an FP8 KV cache.

One caveat about NVFP4: true hardware acceleration, with native FP4 Tensor Cores, requires NVIDIA's Blackwell architecture, meaning the RTX 50 series, RTX PRO Blackwell or GB10. On older NVIDIA cards (Ada Lovelace / RTX 40 series, Ampere / RTX 30 series) or on AMD cards such as the R9700 with ROCm, the format has to be emulated or dequantized to FP8 or BF16 at runtime, for example through Marlin kernels in vLLM. The weights use roughly the same amount of VRAM, but you do not get the full speed gain.

Open or full variant: read the license

Each size comes in two variants.

The full variant has the NB license, an adaptation of Apache 2.0 with some restrictions on use. You may not use the model to reproduce training data, and it may not be used for services that mainly provide access to the licensed press publications the model was trained on.

The open variant has the Gemma license and is trained without the newspaper material from the agreement.

The National Library writes that the full models do slightly better than the open ones on some measurements. Read the license for the variant you choose against how you will actually use it.

Is Borealis better than other models?

It depends on the task.

The National Library's own measurements show that Borealis is strongest where it was built to be strong: Norwegian language, knowledge of Norwegian conditions and following instructions. They also write openly that Borealis does not win every measurement. NorMistral is very strong on some Sámi tasks and on translation, and Gemma is strong on several general metrics.

Our own test from July 2026 says something about the neighbourhood, but not about Borealis itself. We gave five open models five real work tasks in Norwegian, on our own hardware, and let them judge each other blind. None was best at everything. Gemma-4-26B produced the most natural Norwegian prose in the field. Borealis was not in the test, and the panel of judges is not calibrated against human assessment. Note the generation gap too: Borealis builds on Gemma 3, while it is Gemma 4 that was tested, and that Borealis 2 builds on. Read the test of the local models.

So our advice is the same as for any model: test it on your own tasks before you rely on it in production.

How we run Borealis ourselves

We run Borealis on an internal Dell Pro Max with GB10. It is a small desk-side machine built on NVIDIA's GB10 Grace Blackwell chip, with 128 GB of coherent memory shared between CPU and GPU, around 1000 TOPS in FP4, and NVIDIA DGX OS.

Because the memory is unified, a 27B model is nowhere near a challenge. There is ample room for a large context window, or for running several models at once. The box also has ConnectX-7, so two units can be linked for 256 GB if the need grows.

For us the point is that it is small enough to sit in ordinary office premises, and that it has the Blackwell architecture needed to run NVFP4 the way the format is meant to be run. We deliver this kind of hardware fully set up, with models, vLLM and integrations in place, whether it sits with you or is operated by us.

How to get started

  1. Choose a size. Start with 4B or 12B. Move up to 27B if the quality is not good enough.
  2. Choose a variant and read the license. Full or open, see above.
  3. Run the model. Ollama is the easiest: ollama run hf.co/NbAiLab/borealis-12b-gguf. For production with several users, the models are also available for vLLM and Transformers.
  4. Test on your own tasks. Pick one narrow job, for example summarising meeting minutes or drafting replies to customers. Use the same examples every time, and have a subject expert read the answers.
Two ways to run BorealisBoth are valid. The choice is about how many people will use it.Simple startFORMATGGUF, run with OllamaMACHINELaptop, Mac or a single cardFITS WHENOne or a few people want to trythe model on their own tasks.ProductionFORMATNVFP4 or BF16, run with vLLMMACHINEA server, or a GB10 boxFITS WHENMany users, longer contextand uptime requirements.NVFP4 gets full hardware acceleration only on Blackwell. On older cards the format is emulated at runtime.

All the models are gathered at the National Library on Hugging Face.

What to be aware of

The National Library itself says that Borealis can be wrong and make up answers, and that the models should not be used in safety-critical settings without your own evaluation and safeguards.

The smallest sizes suit narrow tasks, not answering everything.

When you run the model yourself, you are also responsible for operations, updates and security.

When does it pay to run the model yourself?

It pays when the data is sensitive, when you want to know exactly where the text is processed, or when the volume is large enough that your own machine becomes cheaper than paying per use.

With us, the solutions run on our own servers in Sandefjord by default. If you have sensitive data, we set it up locally at your premises, so the data never leaves the office. It can also run in Europe or with the provider you choose.

Read more about local and Norwegian AI and about how we choose the model for the task.

Questions and answers

Is Borealis free? Yes. The models are free to download and use within the license. The cost lies in the machine they run on and in operations.

Can we use Borealis commercially? Yes, within the license for the variant you choose. The full variant has the NB license with some restrictions, and the open one has the Gemma license.

What does it take to run Borealis 27B? The compressed version (Q4_K_M) is 16.5 GB. Expect to need a graphics card with 24 GB of memory, or a server.

Where does Borealis run? Where you choose. On a PC, on your own server or with a provider you trust.

What is the difference between Borealis and Borealis 2? Borealis builds on Gemma 3 and was presented in May 2026. Borealis 2 builds on Gemma 4 and is available as a preview for testing.

Sources

  • Sizes, Gemma 3, variants, licenses, own assessments of strengths and weaknesses: ai.nb.no/borealis (read 19.11.2026)
  • The 10 August launch, the Kopinor agreement, NOK 45 million and the Karianne Tung quote: the National Library's press release via NTB, 10.8.2026
  • Q4_K_M file sizes: NbAiLab on Hugging Face (read 19.11.2026)
  • Borealis 2: the model card for borealis2-31b-preview (Gemma 4, release date 28 September 2026, preview)
  • Our own model test, July 2026: No AI model is best at everything
From article to practice

Want to put this to work?

Local and Norwegian AI. By default, the solutions we build run on our own servers in Sandefjord. If your data is sensitive, we set them up on site at your premises.

Read about Local and Norwegian AI Book a free call
Safe and local AI

Read on.

Newsletter

AI in practice, one email a month.

What we've tested, one tip you can use, and what's coming next.

You'll get the Skjld Labs newsletter once a month, and can unsubscribe at any time. Privacy

First step

Thirty minutes. Free.

A no-obligation conversation about what you are facing. We will say so if AI is not the right answer. We are happy to meet you at Langes gate 8 in Sandefjord, or on video.