Offline AI

Let's find the right setup for your computer

Answer two quick questions and we'll show you which instructions to follow and suggest an AI model to try. Don't worry if you're unsure — we'll help you find the answers.

Before you begin: 3 quick checks

1. Your computer runs Windows, macOS or Linux and has permission to install software. If a work computer asks for an admin password you don't have, install on a personal machine instead.
2. Internet access for the initial downloads (the tools and the AI model download over the internet once; after that, answering questions happens on your machine — but see our privacy note below, because parts of these tools can still contact the internet).
3. Memory and disk space. As a rough guide: the model we suggest for 16GB machines downloads about 5GB and installing the two apps takes a few hundred MB. Downloads over a normal home connection take roughly 5–20 minutes; the whole setup, typically 45–90 minutes. These vary — a lot depends on your connection and computer.

Wondering if your computer is powerful enough? The hardware guide explains it in plain English — but you don't need it before continuing. If your computer is 8GB or less, read the gentle warning the selector above will show you and consider the online AI guide instead.

The three tools, in plain English

YOU
 |
 v
KILO CODE
(Inside VS Code)
 |
 v
OLLAMA
 |
 v
LOCAL AI MODEL
(On your computer)

All three tools are installed on the same computer in this tutorial — nothing is being sent to a server for processing.

Step 1 — Install your workspace (VS Code)

Step 2 — Install and test your AI model (Ollama)

Qwen3 8B is the model this page suggests for most beginners; the selector above may have suggested a different (smaller) one based on your RAM. The exact identifier is shown in each command block below — you don't need to remember it.

About the download: the model Qwen3 8B is about 5GB. On a normal connection that's usually 5–20 minutes. It's a one-time download per model — afterwards it works offline (though Ollama itself, like any app, may still make network calls — see our privacy section below).

Download and start the model with this command (the selector's model may be smaller — whatever it suggested, use that one; the block updates automatically):

What happens: first, a progress bar downloads the model (this is the big download — leave it running). Then the prompt changes to >>> — you're now chatting with the model on your machine.

Try it: type Why is the sky blue? Answer in one sentence. and press Enter. A short answer appears in a few seconds (or slower on older computers — that's normal).

Leave the conversation: type /bye and press Enter (this exits chat but Ollama keeps running in the background — that's what you want).

If it doesn't work: ‘model not found’ means a typo — retype the exact command from the block above. ‘Error: could not connect’ means Ollama isn't running — on Windows, restart the Ollama app from the Start menu; on Mac, open the Ollama app again; on Linux, run ollama serve.

Step 3 — Connect your AI assistant (Kilo Code)

1. Open VS Code. Click the Extensions icon — four squares in the left sidebar (or press Ctrl+Shift+X on Windows/Linux, Cmd+Shift+X on Mac).
2. Type Kilo Code in the search box, press Enter.
3. Click Install on the Kilo Code result (its icon is a blue K). What happens: ‘Installing’ shows briefly.
4. A Kilo Code icon appears in the left sidebar (a little rocket). Click it — a panel opens with first-run setup.
5. Connect it to your local model: when the panel asks for a provider, look for a settings gear icon in the Kilo Code panel → open Providers. Choose Ollama from the provider list (it's under ‘Local & Self-Hosted’).
6. Model: in the model field, enter the exact model from step 2: ollama/qwen3:8b (the ‘ollama/’ prefix is required — it tells Kilo Code your model is a local Ollama model). No password or key is needed — it's running on your machine.
7. Context size (worth setting, plain-English version): local models only ‘see’ a limited amount of text at a time. In the same settings, find ‘Context Window Size (num_ctx)’ and set it to 32768 if your machine has 16GB+; leave it on default on smaller/older machines.
8. Test it: type in the Kilo Code chat box: Hello — can you see this project? A reply from your local model appears. You're connected.
9. If it doesn't work: see Stuck? Start here below — especially items 5 and 7.

Full current documentation lives at kilo.ai/docs/ai-providers/ollama — if the interface you see differs from these labels (Kilo updates regularly), trust the official docs and follow its steps.

Your first success: one complete exercise

Before real projects, walk through one simple task end to end. This is the moment the guide exists for.

Stuck? Start here

Alternative: Prefer a more visual way to run AI? Try LM Studio

Beginners should choose Ollama or LM Studio — not both. They do the same job in different ways. Ollama suits the guided three-step path above; LM Studio is more visual, with point-and-click model browsing. If you'd rather click than type, use this section instead of Step 2 and give Kilo Code its address instead.

Practical safety and privacy

• What stays local: your model's answers are generated on your computer — prompts and responses don't go to a cloud for inference. That's the point of this setup.
• What may still touch the internet: initial installs and model downloads (obviously), application updates, telemetry/settings sync features in VS Code and extensions, and optional Kilo Code cloud features. If you need strict offline operation for sensitive work, that's an advanced configuration — verify each tool's settings before relying on it.
• What AI gets wrong: local models are smaller than cloud ones and make mistakes — in code especially. Review every change before accepting it. Kilo Code shows you exactly what it wants to change before it does it — that review step is your safety net.
• Never disable security protections (antivirus, SmartScreen, Gatekeeper) to make an install work, and never expose your local AI server to the internet. If a step seems to require either, stop and seek help instead.