OpenAI
GPT
OpenAI’s language models, which also power ChatGPT and Codex.
Via OpenAI’s services and API.
AI models and tools we work with
The right model for the right task. Not one model for everything.
We work with a broad range of leading AI models, including Qwen and GLM, and assess new models continuously. We choose the model and tool based on the task, the quality, the cost and the requirements for data and hosting. A solution can use cloud services, models on our own infrastructure or a local setup at the customer’s premises.
Ask us what suits youFour things we keep apart.
It makes it easier to see what you are actually buying, and where the data is processed.
The provider
The company that develops the model, for example OpenAI, Anthropic, Google, Alibaba or Z.ai.
The model
The language model or model family itself, such as GPT, Claude, Gemini, Qwen or GLM.
The tool
The product you use, such as ChatGPT, Microsoft 365 Copilot or Codex. A tool is not the same as the model it uses.
The hosting
Where the model runs: with the provider, on our infrastructure or at your premises. An open model can also run in a cloud service.
Model families we work with.
The models are good at different things, and that changes quickly. That is why we test continuously and use whichever suits the task best.
Anthropic
Claude
Anthropic’s language models. Used in the Claude app, in Claude Code and via the API.
Via Anthropic’s services and API.
Gemini
Google’s model family, with variants for different tasks and types of content.
Via Google’s services and API.
xAI
Grok
xAI’s model family. We use Grok alongside GPT and Claude, in agents and automations where it suits the task. Web and X search can be added as separate tools when a task needs fresh sources.
Via xAI’s services and API.
Alibaba
Qwen
Open models in many sizes and variants. In our internal test from July 2026, the Qwen model performed best on some tasks in fact extraction and Norwegian summarisation.
Via a cloud service or on our own infrastructure.
Z.ai
GLM
A model family from Z.ai, designed among other things for coding and agent tasks. Several versions are open. Version and hosting follow the setup of each solution.
Via a cloud service or on our own infrastructure.
Gemma
Google’s open models. In the same internal test, Gemma produced the most natural Norwegian text, according to the judging panel.
Via a cloud service or on our own infrastructure.
NVIDIA
Nemotron
NVIDIA’s open models, designed among other things for multi-step reasoning.
Via a cloud service or on our own infrastructure.
NB
National Library of Norway
Borealis
Norwegian language models from the National Library of Norway, built on Gemma 3, in five sizes. They can be self-hosted, but the variants have different licences with their own terms of use.
On our own infrastructure.
The July 2026 test applies to the models, tasks and setup tested. The answers were judged by a panel of models, not by people, and reasoning mode was turned off. It is not a ranking of the market. Read the method and results.
Tools and agent platforms.
The products that use the models. Each has its own terms for data and hosting.
Microsoft
Microsoft 365 Copilot
AI in Teams, Outlook, Word and SharePoint. The product combines language models with your organisation’s own data and apps, and uses the access rights each employee has. Skjld Labs is a Microsoft partner and handles the setup.
When you already use Microsoft 365.
OpenAI
ChatGPT
OpenAI’s assistant for writing, analysis and questions, in the browser, as an app and for businesses.
For broad use across the organisation.
OpenAI
Codex
OpenAI’s agent for software development. Writes, fixes and tests code on request.
When we build and maintain solutions.
Anthropic
Claude Code
Anthropic’s tool for software development, where the Claude models work directly in the code and the terminal.
When we build and maintain solutions.
Gemini in Google Workspace
Gemini features in Gmail, Docs and Drive. Which features are available depends on the subscription and product.
When you work in Google Workspace.
Nous Research
Hermes Agent
An open-source agent, not a language model. It runs on your own machine or server and can be connected to different models. We use it with the local models on our own infrastructure. Nous Research also makes its own Hermes models, which are something different.
For agents that work in several steps, with tools, memory and scheduled tasks.
Three hosting options. We clarify the data flow first.
Where the data is processed is decided by the service, the agreement and the setup. We set up all three, and can combine them.
Option 1
Cloud services from the provider
Models and tools such as ChatGPT, Claude and Microsoft 365 Copilot, via the provider’s service or API. The quickest way to get started. We set them up and teach you to use them safely.
The dataPlace of processing and data terms depend on the service and the agreement.
We check region and terms against your requirements before choosing a service.
Option 2
Our infrastructure in Sandefjord
The models run on our hardware in Sandefjord, and we handle operations and updates. You don’t need hardware of your own.
The dataModel processing can run on our own infrastructure in Norway. Other data flows are clarified for each solution.
When model processing should take place in Norway, without running it yourselves.
Option 3
A local setup at your premises
We can set up the model and hardware in your environment. The scope of operations and follow-up is agreed for each solution.
The dataModel processing can run in your environment. External integrations, operations, logs and backups are clarified separately.
When the data requires it, for example health, finance or trade secrets.
A shared integration layer makes it easier to switch between supported models. New models are tested for quality, tool use, format, privacy and compatibility before they are put to use. Products such as Microsoft 365 Copilot and Codex are not switched in the same way. Read more about local and Norwegian AI.
How we choose a model.
Three questions decide most of it.
What data will be processed?
We clarify the requirements for privacy and place of processing, and choose the hosting accordingly.
What do you already use?
If you have Microsoft 365 or Google Workspace, the AI already built in is often a good place to start.
What does the task require?
We assess quality, context needs, speed and cost for the task before we choose a model and hosting.
Speech recognition
Norwegian speech to text, tuned for Norwegian.
Our speech recognition builds on open speech models trained on Norwegian, with our own integrations and adaptations. In Skjld Diktat, recognition and language clean-up run on Skjld’s servers.
You say
«Eh, send en mail til kunden om at, eh, leveransen kommer på fredag og at vi sender faktura etterpå.»
You get
Send en e-post til kunden om at leveransen kommer fredag, og at vi sender faktura etterpå.
Illustrative example, in Norwegian
01
Speak
Hold the shortcut key and speak naturally.
02
Recognise
The speech model turns the audio into text. Audio is processed to create the transcription.
03
Clean up
Filler words and punctuation are tidied up, with the aim of preserving the meaning. You can check the text before it is used.
04
Finished text
The text is inserted where the cursor is, in programs that support text insertion.
Storage and deletion of audio, transcriptions and local history follow the chosen solution and settings. Always check names, numbers and technical terms before the text is used.
Where the technology is used today
Sister company
MediVox
MediVox develops solutions for Norwegian speech recognition and documentation workflows, with open speech models and its own integrations and adaptations.
Dictation and documentation workflows.
Skjld Labs · in use
Skjld Diktat
Norwegian dictation that inserts the text where the cursor is, in programs that support text insertion. Speech recognition runs on Skjld’s servers.
Speak instead of typing.
Skjld Labs · open beta
Skjld Cloud
Skjld Cloud supports meeting transcription and further work with meeting notes. The speech engine and place of processing depend on the setup.
Less manual work with meeting notes, with the option to review and edit.
Why we know this.
We don’t recommend models we haven’t tried ourselves.
We use AI in our own work
We use AI tools in our development work and choose our own model setups for the features in our products. Model, data processing and hosting are described for each solution.
We test before we recommend
We put five local models to work on five real Norwegian work tasks and had them assess each other blind. None was best at everything.
We run our own infrastructure
The hardware for the models is at our office in Sandefjord. We know what operations cost, and we tell you when the numbers don’t add up.
Questions and answers
What people ask us.
Which model is best?
None is best at everything. It is decided per task and per dataset, not once for the whole organisation.
Where is our data processed?
It depends on the solution. For cloud services, the service, the agreement and the region decide where data is processed. Model processing can also run on our infrastructure in Norway or in your own environment. We clarify the data flow before we build.
Can we switch models later?
Often. A shared integration layer makes it easier to switch between supported models, but new models are tested before they are put to use. Products such as Microsoft 365 Copilot and Codex are not switched in the same way.
Can you set up a model at our premises?
Yes. Model processing can run in your environment. Hardware, operations, integrations, logs and backups are agreed for each solution.
Is an open model free?
The model weights may be available at no cost, but licence terms vary, and hardware, setup and operations cost money. We include that in the calculation before you decide.
Not sure which model suits you? We will figure it out together in 30 minutes.
Book a free conversation