How is Od1n V5 built?
Od1n V5 is built in three deliberate stages: first a foundation model learns the Norwegian language from a large body of text, then it is polished on the best Norwegian material to sharpen quality, and finally an assistant layer teaches it to behave like a helpful, honest assistant. Facts and language are trained into the model once. Each customer then adapts Od1n at deployment by connecting their own documents through retrieval, rather than paying to train a new model. The result is a small, sovereign Norwegian model that is cheap to adapt and hard to break.
That single sentence hides a lot of engineering choices. Below we tell the story of how Od1n, a product of EZ-Fix AS, goes from raw Norwegian text to an assistant you can actually trust in a business workflow.
Stage one: how does a language model learn Norwegian?
Every language model starts by learning to predict language from a large collection of text. For Od1n, that foundation is built on Norwegian. This is the stage where the model absorbs grammar, vocabulary, tone, and a broad sense of how Norwegian is actually written, including both Bokmål and Nynorsk.
This matters because most large models are overwhelmingly trained on English. They can speak Norwegian, but they think in English and translate. Od1n is built the other way around: Norwegian is the native ground, not an afterthought. That is what "sovereign LLM training" means in practice, a model whose competence starts from your own language and context.
Stage two: what does the final polish add?
A foundation model knows a lot but is rough around the edges. The second stage sharpens it on the best Norwegian material available, raising the quality and consistency of what the model produces. Think of it as moving from "understands Norwegian" to "writes Norwegian well."
By the end of these two stages, language and general knowledge are baked into the model itself. This is a deliberate architectural decision: the expensive, heavy learning happens once, centrally, and does not need to be repeated for every customer.
Stage three: how does Od1n become an assistant?
A capable model is not the same as a helpful assistant. The third stage, the assistant layer, teaches Od1n how to behave: how to follow instructions, how to answer questions from provided documents, how to reason through a task, and critically, how to say "I don't know" instead of inventing an answer.
This is the layer that turns a language engine into something a business can put in front of staff or customers. It also handles Nynorsk properly, which surprisingly few Norwegian-capable models do well.
Why separate facts from behaviour?
Here is the key idea for anyone deciding whether to buy or build with Od1n. Language and general knowledge live inside the model. The specific facts a customer cares about, their contracts, manuals, policies, product data, are supplied at deployment through retrieval, commonly called a RAG language model approach.
Instead of retraining a bespoke model for every client, you connect Od1n to that client's own documents. The model reads from those sources when it answers. This makes adaptation cheap and fast, keeps each customer's data separate, and means updating the knowledge is as simple as updating the documents, not rerunning a training pipeline.
How does Od1n avoid making things up?
Hallucination, confidently stating something false, is the single biggest reason businesses hesitate to trust AI. Od1n is built directly against it. The assistant layer trains the model to answer from the connected documents and to admit when the answer is not there.
An honest AI assistant is one that grounds its answers in your sources, refuses unsafe or unanswerable requests without becoming uselessly over-cautious, and tells you plainly when it is uncertain. That combination, grounded answers plus honest limits, is what makes an AI safe to embed in real workflows rather than just demos.
Is a small model good enough?
Small, focused, and already proven. In internal benchmarks Od1n V5 (3B) beats larger models from Google, Meta and Alibaba on Norwegian, and goes past a model more than twice its size on Norwegian public-administration text. A compact model is also cheaper to run, easier to deploy on your own hardware, and simpler to govern.
For a business, the promise is concrete: an AI you can trust inside your processes, one that sticks to your sources, admits when it doesn't know, stays sovereign and under your control, and is cheap to adapt to new use cases. You can read more about the technology and benchmark, see how Od1n can be used, and learn how Od1n runs on-premise.
The short version: build language and knowledge once, adapt with each customer's own documents at deployment, and train the model to be honest about what it does and doesn't know. That is how Od1n V5 goes from a Norwegian foundation to an assistant you can actually rely on.