Practical local AI guide

How to use Ollama with Markdown notes

A reproducible path from a fresh Ollama install to a useful note summary in Ariv, with model choices, privacy boundaries, and the fixes most setups need.

The short answerInstall Ollama, run ollama pull gemma3:4b, confirm the local API responds at http://localhost:11434, then open Ariv's Settings > AI & Intelligence. Choose Ollama (Local), keep the default URL, select the model, and summarize a non-sensitive test note.

Ollama turns a downloaded language model into a service on your computer. Ariv can use that service for its note summary, Ask Brain, and AI-assisted tag refinement workflows. Your Markdown files remain your notes. Ollama is the selected inference route, not a replacement file format.

Before you start

You need a macOS or Windows computer, Ariv, enough free disk space for a model, and an internet connection for the initial downloads. Ollama's current macOS documentation requires macOS Sonoma or newer. Windows and Linux installers are also available from Ollama.

Model pages list download sizes, but those numbers are not exact runtime memory requirements. Available memory, context length, other running apps, model architecture, and hardware acceleration all affect real performance.

Start small, prove the route, then upgrade.A 4B model is usually the most sensible first test. It lets you validate the connection and workflow before committing storage and memory to a larger model.

1. Install Ollama

Download Ollama from the official quickstart and complete the installer for your operating system. On macOS, Ollama recommends installing the application in the Applications folder.

After installation, open Ollama. The local API is normally available automatically at http://localhost:11434/api.

2. Pull a useful starter model

Open Terminal on macOS or PowerShell on Windows and download one local model. This guide uses Gemma 3 4B because its official Ollama page specifically lists summarization among its strengths and the download is currently about 3.3 GB.

Terminal or PowerShell
ollama pull gemma3:4b
Gemma 3 4Bgemma3:4b · about 3.3 GBA practical starter for summarization and general note work. Its Ollama listing shows a 128K context window.
Qwen 3.5 4Bqwen3.5:4b · about 3.4 GBAn alternative compact model with a longer 256K listed context window. Test both on your own notes before choosing by reputation.

Larger models can improve difficult synthesis, but they also require more storage, memory, and response time. For the first connection test, model size is less important than a successful end-to-end result.

Avoid :cloud tags for a local-only setup.Ollama now offers cloud models that offload work to Ollama's cloud. If local inference is your goal, choose a downloaded local tag such as gemma3:4b, not a tag ending in :cloud. Ollama also documents a local-only mode that disables cloud features.

3. Verify Ollama before involving Ariv

First, confirm that Ollama sees the model:

List installed models
ollama list

Then check the same local model endpoint Ariv uses to populate its model selector:

Check the local API
curl http://localhost:11434/api/tags

You should receive JSON containing a models list. If this command fails, fix Ollama before changing Ariv settings. That keeps the diagnosis simple.

4. Connect Ariv to Ollama

Open Ariv and follow this path:

SettingsAI & IntelligenceAI ProviderOllama (Local)

Keep Ollama URL set to http://localhost:11434 when Ariv and Ollama are on the same computer. Then choose gemma3:4b, or your installed model, under Ollama Model.

Do not paste /api or /api/chat into the URL field.Use the server base URL. Ariv adds the API route required by each request.

5. Test with a deliberately messy note

Create a disposable note with enough substance to summarize. Fictional content is better for the first test because it proves the route without putting a sensitive note at risk while the setup is still new.

ollama-test.md
# Launch thoughts

call sam maybe tuesday, pricing page still says old number,
launch probably september 4 unless onboarding bug is not fixed.
Maya needs screenshots by friday. risk: windows signing cert expires soon.
also should we email the beta group before the public post? decide tomorrow.
Remember the product name is Northstar, not North Star.

Open the note and run Ariv's Summary action. A successful result should produce a short summary plus structured fields such as action items, key dates, and risks when the note contains them. Review the output before appending anything to the source note.

If the result is weak but the request completes, the integration works. Improve quality separately by clarifying the source note, trying a different local model, or moving to a larger model your hardware can run comfortably.

Want the local AI workflow around plain Markdown?

Ariv is free on macOS and Windows. Ollama remains a separate local runtime you control.

Try Ariv free

What stays local, exactly?

With a downloaded local model selected, Ariv sends supported AI requests to the Ollama URL you configured. At the default desktop URL, that service is on the same machine. Ollama performs the model inference there, and your vault continues to exist as Markdown files.

That statement does not automatically cover every optional feature in either product. Ariv updates, telemetry, optional sync, and separately configured services can use a network. Ollama's :cloud models intentionally use its cloud. Privacy depends on the actual model tag, URL, and feature settings you choose.

Local AI is a configuration you can verify: local files, a local server URL, and a downloaded local model.

Troubleshooting the common failures

1Ariv cannot reach Ollama

Open Ollama and run curl http://localhost:11434/api/tags. If that fails, restart Ollama. Keep the Ariv URL at http://localhost:11434 for a same-computer setup.

2The model list is empty

Run ollama list. If no models appear, pull one first. Reopen the model selector after the download completes.

3The first answer is very slow

The model may be loading into memory. Try a second request. If every request remains impractical, use a smaller model and close memory-heavy applications.

4The response is malformed or unhelpful

Try a current instruction-tuned model, shorten the test note, and confirm the model can complete a basic chat in Ollama itself. Connectivity and answer quality are separate problems.

5Ollama is on another computer

localhost always means the device running Ariv. A LAN setup needs the other computer's address and careful Ollama network configuration. Exposing an unauthenticated model server beyond a trusted network is not recommended.

Where to go next

Read the local AI notes hub for the product and privacy model, or compare six local AI note-taking apps and setups if you have not chosen a workflow yet.

Official sources and verification

Product routes and labels were checked against the current Ariv codebase on August 21, 2026. Model listings and software requirements can change, so confirm them on the linked official pages before downloading.

Start with one local model and one honest test note.

Once the full route works, you can decide whether a larger model earns its storage and memory cost.

Download Ariv free