AI Coding Setup: Local LLMs with Cursor, VS Code & Continue.dev (2026)

Last updated: March 31, 2026. Tested with Cursor 0.44, Continue.dev 0.9, and Ollama 0.6.x on Ubuntu 24.04 and macOS 15.

Why Local AI Coding?

Cloud AI coding tools cost $20–$40/month per seat, send your code to remote servers, and throttle you with rate limits when you need them most. A local AI coding setup costs $0/month after hardware, sends zero code to any cloud, and has no rate limits.

In 2026, local coding models are genuinely good. DeepSeek-Coder-V2 and Qwen2.5-Coder match or exceed GPT-4-level coding quality on many benchmarks when running locally at the right model size. This guide sets up a complete local AI coding environment from scratch.

What You Need

Component	What It Does	Cost
GPU with 8GB+ VRAM	Runs the AI model locally	Hardware you (may) already own
Ollama	Runs and serves local models via API	Free
IDE plugin	Connects your editor to Ollama	Free (or Cursor's base tier)

For hardware recommendations, see our GPU buyer's guide. The minimum for a useful experience is an 8GB GPU like the RTX 3060. For the best local coding assistant, an RTX 4090 or RTX 3090 running a 32B model is compelling. Even a Mac Mini M4 Pro runs 7B–14B coding models silently and well.

Step 1: Install Ollama and Pull a Coding Model

Install Ollama (full guide: Ollama setup guide):

# Linux
curl -fsSL https://ollama.com/install.sh | sh

# macOS
brew install ollama

Pull a coding model. Here are our recommendations by VRAM:

# 8GB VRAM — fast 7B coder
ollama pull qwen2.5-coder:7b

# 16GB VRAM — best mid-range coding model
ollama pull deepseek-coder-v2:16b

# 24GB VRAM — best single-GPU coding experience
ollama pull qwen2.5-coder:32b

# Any VRAM — reasoning-focused coding (slower but smarter)
ollama pull deepseek-r1:14b

Verify the model runs:

ollama run qwen2.5-coder:7b "Write a Python function to binary search a sorted list"

Best Local Coding Models in 2026

Model	VRAM	Speed	Best For	Ollama Command
Qwen2.5-Coder 7B	5GB	~90 t/s (4090)	Fast completion, syntax help	ollama pull qwen2.5-coder:7b
DeepSeek-Coder-V2 16B	10GB	~50 t/s (4090)	Code generation, multi-file	ollama pull deepseek-coder-v2:16b
Qwen2.5-Coder 32B	20GB	~35 t/s (4090)	Best single-GPU coding quality	ollama pull qwen2.5-coder:32b
DeepSeek-R1 14B	9GB	~35 t/s	Debugging, reasoning through bugs	ollama pull deepseek-r1:14b
CodeLlama 34B	22GB	~30 t/s	Code completion, fill-in-middle	ollama pull codellama:34b

Option A: Cursor with Local Ollama

Cursor is the most popular AI-native code editor. It ships with Claude and GPT-4 integration built in, but also supports custom API endpoints — which means you can point it at your local Ollama instance.

Setup

1. Download and install Cursor (free tier available)
2. Open Settings → Models
3. Under "OpenAI API Key", enter: ollama
4. Under "OpenAI Base URL", enter: http://localhost:11434/v1
5. Add your model name (e.g., qwen2.5-coder:32b) to the model list
6. Select your local model from the model picker in the chat panel

You now have Cursor's editing interface — inline edits, multi-file context, the agent mode — running against your local GPU instead of OpenAI's servers.

What Works (and What Doesn't)

Works fully locally: Chat panel, inline edits (Cmd+K), code generation, refactoring requests.

Still uses cloud (Cursor's servers): Cursor's proprietary indexing and codebase search features. If your code is sensitive and you want zero cloud exposure, use Continue.dev instead.

Option B: VS Code + Continue.dev (Fully Local)

Continue.dev is an open-source VS Code extension designed specifically for local LLM integration. Everything runs locally — no telemetry, no cloud, no code leaving your machine.

Setup

1. Install the Continue extension in VS Code
2. Click the Continue icon in the sidebar
3. Open ~/.continue/config.json and add:

{
  "models": [
    {
      "title": "Qwen2.5-Coder 32B (Local)",
      "provider": "ollama",
      "model": "qwen2.5-coder:32b",
      "apiBase": "http://localhost:11434"
    },
    {
      "title": "DeepSeek-R1 14B (Reasoning)",
      "provider": "ollama",
      "model": "deepseek-r1:14b",
      "apiBase": "http://localhost:11434"
    }
  ],
  "tabAutocompleteModel": {
    "title": "Qwen2.5-Coder 7B (Autocomplete)",
    "provider": "ollama",
    "model": "qwen2.5-coder:7b",
    "apiBase": "http://localhost:11434"
  }
}

This config sets up two chat models (a capable coder and a reasoning model for debugging) plus a fast autocomplete model. The 7B model handles real-time completion without noticeable latency.

Key Features

• Inline autocomplete: Press Tab for AI-suggested completions as you type
• Chat with your code: Cmd+L to open chat with selected code as context
• Edit mode: Cmd+I to request inline edits on selected code
• @-mentions: Reference files, docs, and functions directly in the chat
• Zero telemetry: Everything stays local, configurable in settings

Access Your AI Coding Assistant From Any Device

If you have a desktop AI server (see our home AI server guide), you can access it from your laptop anywhere on your network:

On the server, configure Ollama to listen on all interfaces:

# Edit Ollama service
sudo systemctl edit ollama
# Add: Environment="OLLAMA_HOST=0.0.0.0"
sudo systemctl restart ollama

In Continue's config on your laptop, change apiBase:

"apiBase": "http://[server-ip]:11434"

Your laptop's keyboard and monitor drive the coding session, while the heavy model inference runs on your desktop GPU. The laptop stays cool and fast; the GPU server does the work.

Real-World Performance Comparison

Setup	Autocomplete Speed	Code Gen Quality	Monthly Cost	Privacy
GitHub Copilot (cloud)	Fast	Good	$19/month	Code sent to Microsoft
Cursor Pro (cloud)	Fast	Excellent (GPT-4)	$20/month	Code sent to Cursor/OpenAI
Local 7B (RTX 3060)	Fast (~80 t/s)	Decent	$0/month	Fully local
Local 16B (RTX 4060 Ti)	Medium (~45 t/s)	Good	$0/month	Fully local
Local 32B (RTX 4090)	Medium (~35 t/s)	Excellent	$0/month	Fully local

For developers working on proprietary code, the privacy argument alone often justifies local setup. For freelancers and solo developers, the cost savings compound fast — $20/month is $240/year, every year, indefinitely.

Getting Started

The fastest path to a working local AI coding setup:

1. Install Ollama: curl -fsSL https://ollama.com/install.sh | sh
2. Pull a coding model: ollama pull qwen2.5-coder:7b (or larger for better quality)
3. Install Continue.dev in VS Code and point it at localhost:11434
4. Open a file you are working on and press Cmd+L to start chatting with your code

The whole setup takes under 10 minutes. Once it is running, you will never go back to cloud-only coding tools for anything proprietary or sensitive — and you will appreciate that the bill stays at $0.