Hands-on guide · VS Code · Ollama + LM Studio

How to use VS Code with Local Models

Run an LLM on your own machine and wire it into VS Code — private, offline, and free. Two backends, one bridge, about ten minutes.

Goal: local AI in your editor Model: qwen3:4b Copy-ready commands

Option A

Ollama

CLI-first and the fastest way to start. One command pulls a model; it serves on a local port.

Option B

LM Studio

A point-and-click app. Browse and download models in a GUI, then flip on a local server.

SETUP Before you start

What you need

VS Code — any recent version.
~3 GB free disk for the model download.
A GPU is optional. The model runs fine on CPU alone; a GPU just makes it faster.

The number that matters most is RAM. qwen3:4b at Q4 needs only about 2.5 GB for the weights, so it fits comfortably on modest machines:

Tier	RAM	What it runs
Minimum	8 GB	qwen3:4b on CPU, thinking mode off
Recommended	16 GB	qwen3:4b with room for longer context
Optional GPU	4 GB+ VRAM	Noticeably faster — not required

Long context and reasoning mode are what push a 4B model toward needing 16 GB+. For quick tasks, 8 GB is enough.

HOW IT WORKS The one bridge

The Continue bridge

VS Code only talks to GitHub Copilot out of the box. To reach a local model server, you need a connector — Continue, a free extension that plugs any local backend into the editor. Install it once; both options below use it.

vs code·Extensions panel

# search "Continue" in Extensions, or run:
code --install-extension Continue.continue

Heads-up: Continue is now maintained under Cursor and the open-source repo is frozen at its final release — but it still installs and the local-model path works fully. Cline is the active alternative if you prefer.

OPTION A CLI-first

Run it with Ollama

1Install Ollama

terminal·install

# macOS / Linux
curl -fsSL https://ollama.com/install.sh | sh
# Windows: download the installer from ollama.com

2Pull the model

terminal·pull

ollama pull qwen3:4b

3Test it

terminal·run

ollama run qwen3:4b          # chat here, /bye to exit

Ollama now serves on http://localhost:11434 — that's what Continue connects to.

4Add it to Continue

Open ~/.continue/config.yaml (%USERPROFILE%\.continue\config.yaml on Windows) and add:

yaml·config.yaml

models:
  - name: Qwen3 4B (Ollama)
    provider: ollama
    model: qwen3:4b
    apiBase: http://localhost:11434
    roles: [chat, edit]

OPTION B GUI

Run it with LM Studio

1Install & download the model

Install LM Studio from lmstudio.ai.
In the Discover tab, search qwen3 4b and download the Q4_K_M build.

2Start the local server

Open the Developer / Local Server tab and click Start Server.
It listens on http://localhost:1234/v1 — an OpenAI-compatible endpoint. Note the model ID it shows.

3Point Continue at it

yaml·config.yaml

models:
  - name: Qwen3 4B (LM Studio)
    provider: lmstudio
    model: qwen3-4b            # match the ID in LM Studio
    apiBase: http://localhost:1234/v1
    roles: [chat, edit]

Keep LM Studio's server toggled on while you work — if the server stops, Continue can't reach the model.

DAILY USE In the editor

Use it in VS Code

Click the Continue icon in the sidebar and pick your model from the dropdown.
Ask questions in the chat panel, or reference files with @filename.
Highlight code and press Ctrl/Cmd + I to edit it in place.

✓

That's the whole loop. A private Copilot running on your own hardware — your code never leaves the machine, and there's no API bill.

IF IT BREAKS Quick fixes

If it breaks

“Model not found” → the model: value must match exactly: ollama list for Ollama, or the loaded ID for LM Studio.
No reply → the backend isn't running. Start Ollama (tray icon) or LM Studio's server.
Sluggish or out of memory → turn thinking mode off, shorten the context window, or close other heavy apps.