Skip to content

Installing Ollama on Windows 11

Audience: anyone who wants the AI narration to run locally. Time: about 15 minutes plus the model download. You need: Windows 11 and disk space for the model — no admin rights, no account, no API key. At the end: --narrate drafts narrations on your machine, and no data ever leaves it.

This guide is also available as a separate PDF: EN / DE.

1. Why local?

Ollama runs language models entirely on your own machine. For the situation report's AI narration this means: the delta briefing never leaves the system, and you need no corporate approval for external services. This path shows up as deployment_class "local" in the AI banner and in the operator-evidence log (llm_audit.jsonl).

2. Prerequisites

  • Windows 11, 64-bit; no administrator rights needed (Ollama installs into the user profile).
  • Disk space: ~4 GB for Ollama itself plus the model — mistral-nemo downloads about 7 GB, mistral about 4 GB.
  • Memory: 16 GB recommended for mistral-nemo (12B); with 8 GB pick the smaller mistral (7B).
  • Without a GPU everything runs on the CPU — a narration then takes up to a minute. An NVIDIA or AMD GPU is used automatically.

3. Install

  1. Open ollama.com/download/windows and download OllamaSetup.exe.
  2. Double-click the file and follow the wizard — there are no choices to make.
  3. Ollama then runs in the background; the llama icon appears in the taskbar's notification area, and it starts automatically after a reboot.

Verify the installation (Windows key, type powershell):

ollama --version

4. Pull the model

The narration's default model is mistral-nemo:

ollama pull mistral-nemo

On an 8 GB machine, use ollama pull mistral instead. Quick test — the first answer takes longest because the model is loaded into memory (leave the chat with /bye):

ollama run mistral-nemo "Answer in one sentence: what is a situation report?"

5. Wiring check with SituationReport

python -m llm providers
python -m llm test

providers lists the discovered AI backends (ollama, claude, mock); test runs one labeled sample narration through Ollama — if it prints a draft with the AI banner, everything is wired.

6. Use it

CLI — --narrate adds the labeled Narration (Entwurf) section to the delta briefing; the numbers guard discards any invented figure, and the evidence record lands in llm_audit.jsonl next to the output:

python -m portfolio --delta prev.json now.json --narrate --output delta.html --browser

GUI — in the Solutions & Portfolios window tick AI narration (draft), pick the provider next to it, then run Delta briefing … as usual.

7. Troubleshooting

Symptom Fix
Python was not found (points to the Microsoft Store) Windows redirects python to a Store placeholder. Run the commands in the repository folder and use .venv\Scripts\python.exe there (or activate once: .venv\Scripts\Activate.ps1); alternatively py -m llm test, or disable the app execution aliases python.exe/python3.exe under Settings → Apps → Advanced app settings.
Could not reach Ollama / connection refused Ollama is not running — start it from the Start menu, wait for the llama icon.
Ollama does not know model … Run ollama pull mistral-nemo (section 4).
Answers are slow Up to a minute on pure CPU is normal; --llm-model mistral roughly halves the wait.
Low space on C: setx OLLAMA_MODELS "D:\ollama-models", then restart Ollama.
Stop Ollama Right-click the llama icon → Quit Ollama.

8. Privacy in one sentence

Ollama listens on your machine only (localhost:11434); neither briefing nor narration ever leaves the system — and the difference from the external path (Claude API) stays visible at all times via the deployment_class in banner and audit.