Run AI locally: for those who don't want to share data with anyone at all

This entire site is about where your data ends up. Local AI is the endpoint: the model runs on your own computer. Nothing leaves it. You don’t have to audit suppliers. You don’t have to worry about jurisdiction. You don’t have to pay subscriptions. For lawyers, healthcare providers, municipalities and anyone with data that must not leak, this option is simple.

One caveat right away: local models are not as sharp as the largest cloud models. Swedish language quality also varies between models. But for drafts, summaries, coding help and transcription, the gap has narrowed a lot. The price is zero SEK, forever. That’s hard to beat.

Choose your path in

• Easiest (graphical): LM Studio. Install, pick a model from the catalog, chat. The app shows which models your computer can handle. Free even for businesses since 2025. Paid plans only exist for team features such as company-wide login.
• The standard: Ollama. Type ollama run mistral in the terminal, done. Biggest ecosystem, free and open source.
• Open source with a UI: Jan, with LM Studio’s ease of use, free even for businesses.
• Speech to text: whisper.cpp, which transcribes Swedish surprisingly well, completely offline. It is the safest way there is to transcribe a sensitive meeting.

Model tips for Swedish: start with the Mistral or Gemma family in quantized format (Q4). They give good Swedish for the size. Mistral also has European origins.

How much computer do you need? (requirement levels, not model numbers)

Two practical truths:

• Mac is the shortcut. Apple’s unified memory lets even a laptop run large models.
• On PC, it’s the amount of VRAM that counts, not the GPU generation. An older card with 24 GB beats a new one with 12 GB for this purpose.

Specific machines and graphics cards from Swedish resellers are in the buying guide AI computer for the company.

For businesses: local AI as real operations

If local models are to be used for real in an organization, with multiple users, quotas and traceability, operational tools are needed around the model. One open alternative is rs-llmctl. It runs local models as a server with an OpenAI-compatible API, resource budgets and audit logs. (Transparency: the tool is developed by aivalet’s founder. We link to it because it is free and open source, not because we earn anything from it.) For simpler needs, Ollama’s built-in API goes a long way.

When local AI is right, and when it is wrong

Right for: sensitive personal data, client material, patient-facing notes, confidentiality, transcription of internal meetings, or when subscription costs need to go.

Wrong for: when you need absolute best quality (the largest cloud models are still better), real-time collaboration or mobile use.

Many choose a mix: Mistral in the cloud for everyday use, local for the sensitive stuff.

All local tools are in the catalog under Local AI. They are labeled with the text “the data stays with you”. It is the only category where the CLOUD Act column says not applicable.