Continue.dev + Ollama Setup: VS Code config.yaml Example (2026)

name: Local AI Assistant
schema: v1
models:
  - name: Llama 3.1 8B          # chat + edit
    provider: ollama
    model: llama3.1:8b
    apiBase: http://localhost:11434
    roles: [chat, edit, apply]
  - name: Qwen Coder 1.5B       # tab autocomplete
    provider: ollama
    model: qwen2.5-coder:1.5b
    apiBase: http://localhost:11434
    roles: [autocomplete]
  - name: Nomic Embed           # @codebase search
    provider: ollama
    model: nomic-embed-text
    apiBase: http://localhost:11434
    roles: [embed]

What is Continue.dev?

Continue.dev is a leading open-source AI coding assistant (Apache 2.0), offering a free alternative to GitHub Copilot that runs entirely on your machine with local models.

Key Statistics

Why Continue + Ollama?

1. Free forever - No $10-20/month subscription
2. 100% private - Code never leaves your machine
3. Fully customizable - Any model, any workflow
4. Open source - Audit, modify, contribute
5. Enterprise-ready - Used by Siemens, Morningstar

Installation

Step 1: Install Ollama

macOS:

brew install ollama

Linux:

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

Windows: Download from ollama.com and run the installer.

Step 2: Pull Required Models

# Start Ollama
ollama serve

# Autocomplete model (fast, small)
ollama pull qwen2.5-coder:1.5b

# Chat model (quality, reasoning)
ollama pull llama3.1:8b

# Embeddings for codebase search
ollama pull nomic-embed-text

# Verify
ollama list

Step 3: Install Continue Extension

VS Code:

1. Open Extensions (Cmd/Ctrl + Shift + X)
2. Search "Continue"
3. Click Install

JetBrains:

1. Settings → Plugins → Marketplace
2. Search "Continue"
3. Install and restart

Step 4: Configure Continue

Continue reads configuration from:

• macOS/Linux: ~/.continue/config.yaml
• Windows: %USERPROFILE%\.continue\config.yaml

Complete Configuration Guide

Basic config.yaml

name: Local AI Assistant
version: 1.0.0
schema: v1

models:
  # Chat and reasoning (quality model)
  - name: Llama 3.1 8B
    provider: ollama
    model: llama3.1:8b
    apiBase: http://localhost:11434
    roles:
      - chat
      - edit
      - apply
    defaultCompletionOptions:
      temperature: 0.7
      contextLength: 8192

  # Tab autocomplete (fast model)
  - name: Qwen Coder 1.5B
    provider: ollama
    model: qwen2.5-coder:1.5b
    roles:
      - autocomplete
    autocompleteOptions:
      debounceDelay: 250
      maxPromptTokens: 1024
      multilineCompletions: auto

  # Embeddings for @codebase
  - name: Nomic Embed
    provider: ollama
    model: nomic-embed-text
    roles:
      - embed

# Context providers
context:
  - provider: code
  - provider: docs
  - provider: diff
  - provider: terminal
  - provider: folder
  - provider: codebase

# Coding rules
rules:
  - Give concise, focused responses
  - Follow existing code style
  - Prefer TypeScript over JavaScript

Advanced Configuration (24GB+ VRAM)

name: Power User Config
version: 1.0.0
schema: v1

models:
  # Primary chat / edit model (recommended)
  - name: Qwen3 Coder 30B
    provider: ollama
    model: qwen3-coder:30b
    apiBase: http://localhost:11434
    roles:
      - chat
      - edit
      - apply
    capabilities:
      - tool_use  # Enable agent mode
    defaultCompletionOptions:
      temperature: 0.7
      contextLength: 16384
      top_p: 0.9

  # Fast autocomplete
  - name: StarCoder 3B
    provider: ollama
    model: starcoder2:3b
    roles:
      - autocomplete
    autocompleteOptions:
      debounceDelay: 200
      maxPromptTokens: 2048
      multilineCompletions: auto

  # Embeddings
  - name: Nomic Embed
    provider: ollama
    model: nomic-embed-text
    roles:
      - embed

# Custom slash commands
prompts:
  - name: test
    description: Generate unit tests
    prompt: |
      Write comprehensive unit tests for this code.
      Use Jest/Vitest. Cover edge cases.

  - name: refactor
    description: Refactor for readability
    prompt: |
      Refactor this code for better readability.
      Explain your changes.

  - name: review
    description: Code review
    prompt: |
      Review this code for:
      - Bugs and edge cases
      - Performance issues
      - Security concerns
      - Code style
      Provide actionable feedback.

# MCP servers for extended functionality
mcpServers:
  - name: filesystem
    command: npx
    args:
      - "-y"
      - "@anthropic/mcp-filesystem-server"
      - "/path/to/allowed/directory"

Autodetect Models (Simplest)

name: Simple Config
schema: v1

models:
  - name: Autodetect
    provider: ollama
    model: AUTODETECT
    roles:
      - chat
      - edit
      - autocomplete

Best Models by Hardware

4-8GB VRAM (RTX 3060, M1/M2 8GB)

12-16GB VRAM (RTX 4070, M2 Pro)

24GB+ VRAM (RTX 4090, M3 Max)

Updated June 2026: qwen3-coder:30b (30B Mixture-of-Experts, ~3.3B active params, 256K context) is now the recommended local chat/edit model for Continue — it replaces the older qwen2.5-coder line. The MoE design keeps inference fast despite the 30B size, and the q4_K_M quant (~19GB) fits comfortably on a 24GB GPU. deepseek-r1:32b remains a strong alternative when you want explicit chain-of-thought reasoning. If you're picking a chat/edit model, our roundup of the best local AI models for programming and the best 14B coding models compare quality, speed, and VRAM head-to-head.

Which Ollama autocomplete model is best for config.yaml?

Autocomplete is the role most people get wrong, because it isn't about raw model quality — it's about latency and fill-in-the-middle (FIM) support. Tab completion has to respond in well under ~500ms or it feels broken, which rules out anything larger than roughly 3B parameters on consumer hardware. It also has to be an FIM-trained model: autocomplete feeds the model the code before and after your cursor and asks it to predict the middle, which a plain chat model cannot do well.

Rule of thumb: stay at 1.5B–3B for the autocomplete role and put your big model on chat/edit. Bigger is not better here — a 7B autocomplete model adds latency without a meaningful quality gain for inline suggestions, and many users report it actually feeling worse because the cursor stalls.

Why does my Continue autocomplete produce garbage or syntax errors?

This is the single most common config.yaml mistake, and it's almost always a fill-in-the-middle templating problem, not a model-quality problem. If autocomplete spits out broken syntax, duplicated lines, or completions that ignore the code after your cursor, check these in order:

1. You picked a non-FIM model. Generic chat models (llama3.1:8b, mistral, most "instruct" general models) don't understand the <|fim_prefix|> / <|fim_suffix|> / <|fim_middle|> tokens autocomplete relies on. Use a code model that ships FIM: qwen2.5-coder, starcoder2, or deepseek-coder.
2. Wrong tag for the role. On Ollama, qwen2.5-coder:1.5b already resolves to a FIM-capable build that works for tab autocomplete, while you'll generally want an instruct model for the chat role. Don't reuse one chat model for both roles and expect clean completions.
3. The template wasn't auto-detected. Continue normally infers the FIM prompt template from the model name. If you renamed the model or are proxying through vLLM, the wrong chat template can get applied and the <|fim_prefix|> token ends up missing — producing exactly this garbage output. Pulling the model directly through Ollama (ollama pull qwen2.5-coder:1.5b) avoids this.

For a head-to-head of completion engines, Cline + Ollama takes a different, more agent-first approach than Continue's inline tab autocomplete — worth comparing if FIM tuning keeps fighting you.

How do I connect Continue to a remote or networked Ollama server?

A common 2026 setup is running Ollama on a beefy GPU box (or a home server / VM) and coding from a thin laptop. Continue supports this with one field — point apiBase at the remote host instead of localhost:

models:
  - name: Remote Qwen Coder
    provider: ollama
    model: qwen3-coder:30b
    apiBase: http://192.168.1.136:11434   # remote GPU box IP
    roles:
      - chat
      - edit
      - apply

On the server side you must tell Ollama to listen on all interfaces (by default it only binds to localhost):

# Linux/macOS — expose Ollama on the network
OLLAMA_HOST=0.0.0.0:11434 ollama serve

# Persist it as a service env var instead (Linux systemd):
#   Environment="OLLAMA_HOST=0.0.0.0:11434"

Checklist for remote setups:

• Open the port. Allow inbound 11434 on the server's firewall (and your VPN/subnet only — don't expose Ollama to the open internet without auth in front of it).
• Set CORS if needed. OLLAMA_ORIGINS=* avoids origin-blocked requests from the extension.
• Verify reachability first: curl http://192.168.1.136:11434/version from the laptop before editing config.yaml, so you know whether a failure is network or config.
• Use a VPN/SSH tunnel for anything beyond your LAN. Ollama has no built-in authentication, so a tunnel (e.g. ssh -L 11434:localhost:11434 user@server) is the safe way to reach a cloud GPU box — then keep apiBase: http://localhost:11434 on the client.

Key Features

Tab Autocomplete

Press Tab to accept inline suggestions. Works in any file type.

Config options:

autocompleteOptions:
  debounceDelay: 250      # ms before triggering
  maxPromptTokens: 1024   # context for model
  multilineCompletions: auto
  onlyMyCode: true        # ignore node_modules etc.

Tip: Specialized code models (qwen2.5-coder, starcoder2) outperform GPT-4 for autocomplete.

Chat Interface

• VS Code: Cmd/Ctrl + L
• JetBrains: Cmd/Ctrl + J

Select code and ask questions. Add context with @ mentions:

• @codebase - Search entire project
• @file - Reference specific file
• @folder - Include directory
• @docs - Documentation context

Edit Mode

Select code → Press Cmd/Ctrl + I → Describe changes → Apply

Continue modifies code while preserving formatting and style.

Agent Mode

Enable with capabilities: [tool_use] for autonomous multi-step tasks:

• Read and write files
• Run terminal commands
• Search codebase
• Make multiple changes

models:
  - name: Agent Model
    provider: ollama
    model: llama3.1:8b
    capabilities:
      - tool_use

Custom Slash Commands

Create shortcuts for common tasks:

prompts:
  - name: doc
    description: Add documentation
    prompt: Add comprehensive JSDoc/docstring to this code.

  - name: optimize
    description: Optimize performance
    prompt: Suggest performance optimizations for this code.

Use with /doc or /optimize in chat.

Performance Optimization

Reduce Autocomplete Latency

1. Use small models (1.5B-3B parameters)
2. Increase debounce delay:

   autocompleteOptions:
     debounceDelay: 350
   

3. Disable thinking for Qwen3:

   requestOptions:
     extraBodyProperties:
       think: false
   

Verify GPU Acceleration

# Check GPU usage
ollama ps

# Should show GPU layers loaded
NAME              SIZE    PROCESSOR
llama3.1:8b       4.7GB   100% GPU

Debug Issues

# Restart Ollama with debug logging
pkill ollama
OLLAMA_DEBUG=1 ollama serve

# Check Continue logs
cat ~/.continue/logs/core.log

Continue vs Alternatives

When to Choose Continue

• Privacy-critical projects - No code leaves your machine
• Cost-conscious teams - Save $120-240/year per developer
• Custom workflows - Build exactly what your team needs
• Open source preference - Full transparency and control

When to Choose Copilot

• Quick setup priority - Works out of the box
• Enterprise requirements - SSO, compliance features
• Multi-IDE teams - Wider IDE support
• Training data quality - GitHub's massive codebase

Troubleshooting

"Cannot connect to Ollama"

# 1. Ensure Ollama is running
ollama serve

# 2. Verify port
curl http://localhost:11434/version

# 3. Check config.yaml apiBase
apiBase: http://localhost:11434

Models Not Loading

# Pull models first
ollama pull qwen2.5-coder:1.5b
ollama list  # Verify installed

# Restart Continue
# VS Code: Cmd/Ctrl + Shift + P → "Continue: Reload"

Slow Performance

1. Switch to smaller autocomplete model
2. Increase debounceDelay
3. Reduce contextLength
4. Check GPU usage: ollama ps

Config Not Applied

# Validate YAML syntax
python -c "import yaml; yaml.safe_load(open('~/.continue/config.yaml'))"

# Restart VS Code after changes

MCP Integration

Extend Continue with Model Context Protocol servers:

mcpServers:
  # Database access
  - name: sqlite
    command: npx
    args: ["-y", "mcp-sqlite", "/path/to/db.sqlite"]

  # GitHub integration
  - name: github
    command: uvx
    args: [mcp-server-github]
    env:
      GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}

  # File system access
  - name: filesystem
    command: npx
    args: ["-y", "@anthropic/mcp-filesystem-server", "/allowed/path"]

Use in chat: Type @ → Select "MCP" → Choose resource.

MCP works the same whether the model behind it is cloud or local — for a deeper walkthrough of wiring MCP tools to local models, see our Ollama MCP integration guide.

Key Takeaways

1. Continue + Ollama = Free Copilot alternative with full privacy
2. Use small models for autocomplete (qwen2.5-coder:1.5b / starcoder2:3b) and a large coder for chat — qwen3-coder:30b is the current recommendation
3. nomic-embed-text enables powerful codebase search
4. Agent mode requires capabilities: [tool_use] and 8B+ models
5. Custom slash commands automate your team's workflows
6. MCP servers extend Continue's capabilities infinitely
7. 34,000+ stars prove the community trust (Continue was acquired by Cursor in June 2026; the final v2.0.0 release still runs locally with Ollama)

Next Steps

1. Compare local AI tools for model management
2. Explore AI coding agents for autonomous development
3. Compare Cursor vs Copilot vs Claude Code for alternatives
4. Browse the best AI coding tools ranked side by side
5. Check VRAM requirements for model sizing
6. Learn about MCP servers for tool integration

Continue.dev with Ollama delivers a professional AI coding experience without monthly subscriptions or privacy compromises. Whether you're a solo developer seeking GitHub Copilot features for free, or an enterprise team requiring local deployment for compliance, Continue provides the flexibility and performance to transform your coding workflow.