Windsurf Cascade Local Model — Step-by-Step Ollama Setup Guide

Introduction: The Case for Local AI

In the early days of AI-assisted development, cloud-hosted Large Language Models (LLMs) were the only viable option. Developers had to send their code across the internet to servers owned by OpenAI or Anthropic to receive auto-completions or code suggestions.

By 2026, the developer landscape has fundamentally changed. Open-source coding models have advanced to the point where they can compete directly with proprietary cloud models on standard software engineering tasks.

For many developers, running models locally is no longer just an experiment—it is a production requirement. Local AI offers three massive advantages:

• Absolute Data Privacy: Your proprietary source code, database schemas, and configuration keys never leave your physical machine, ensuring 100% compliance with corporate security policies.
• Zero Latency & Offline Capability: Local models execute instantly without network lag, and they continue to function perfectly even when you are on a flight or in an area with a poor internet connection.
• No Subscription Fees: Running open-source models on your own hardware eliminates the monthly seat fees and token-usage markups associated with cloud providers.

In this guide, we will walk through how to configure Codeium's Windsurf IDE and its powerful Cascade agent to run completely locally using Ollama.

What is Windsurf Cascade?

Windsurf is a next-generation visual IDE developed by Codeium. Built as a highly optimized fork of VS Code, Windsurf is designed from the ground up to support "agentic" developer workflows.

The core of Windsurf's AI capability is Cascade.

Unlike basic autocomplete extensions that suggest single lines of code, Cascade acts as a semi-autonomous coding agent. It can analyze your entire repository, index your local files, write multi-file edits, and execute terminal commands.

While Cascade defaults to Codeium's high-performance cloud models, it also provides a highly flexible configuration panel that allows developers to route its reasoning engine to any custom, OpenAI-compatible API endpoint—including a local server running on your machine.

What is Ollama?

Ollama is an open-source, lightweight tool that allows you to run Large Language Models locally on your macOS, Windows, or Linux machine.

Ollama bundles model weights, configurations, and a highly optimized inference engine into a single, easy-to-use background service. It automatically handles GPU acceleration (leveraging Apple Silicon Unified Memory or NVIDIA CUDA cores) and exposes a local HTTP server that mimics the standard OpenAI API structure.

This standardized API endpoint is what allows external tools like Windsurf Cascade to communicate with your local models seamlessly.

Hardware & Software Prerequisites

Running LLMs locally requires significant computational resources. Ensure your machine meets the following minimum specifications:

Minimum Specs (7B Models)

• • CPU: Apple M1/M2/M3 or Intel Core i7 / AMD Ryzen 7
• • RAM: 16GB Unified Memory or System RAM
• • GPU: Integrated Apple GPU or NVIDIA GTX 1080 (6GB VRAM)

Recommended Specs (14B+ Models)

• • CPU: Apple Max/Ultra series or Intel Core i9 / AMD Ryzen 9
• • RAM: 32GB+ Unified Memory or System RAM
• • GPU: Apple M-Series Max GPU or NVIDIA RTX 4070+ (12GB+ VRAM)

Step-by-Step Ollama Setup Guide

Let's install Ollama and pull a high-performance model optimized for code generation.

Step 1: Download and Install Ollama

Go to the official Ollama website and download the installer for your operating system. Run the installer and let it configure the background service.

Step 2: Pull a Coding Model

Open your terminal and pull a model. We highly recommend DeepSeek-Coder (6.7B) or Qwen2.5-Coder (7B) for the best balance of speed and coding capability on consumer hardware:

# Pull the DeepSeek Coder 6.7B model
ollama pull deepseek-coder:6.7b-instruct

# Or pull the Qwen 2.5 Coder 7B model
ollama pull qwen2.5-coder:7b-instruct

Step 3: Verify the Local Server

Ollama automatically runs an HTTP server on port 11434. You can verify it is active by opening your browser and visiting http://localhost:11434. You should see the message:

Ollama is running

Configuring Windsurf Cascade to Use Ollama

Now that Ollama is running, let's configure Windsurf to route Cascade's reasoning engine to your local endpoint.

Step 1: Open Windsurf Settings

Launch Windsurf, open the command palette (Ctrl+Shift+P on Windows/Linux, Cmd+Shift+P on macOS), and search for "Windsurf: Configure Custom API Endpoint" or open your global settings.json file.

Step 2: Add the Ollama Endpoint Configuration

Add the following configuration block to your settings.json file to map Ollama's OpenAI-compatible endpoint:

{
  "windsurf.customApiEndpoints": [
    {
      "name": "Local Ollama",
      "url": "http://localhost:11434/v1",
      "key": "ollama",
      "models": [
        "deepseek-coder:6.7b-instruct",
        "qwen2.5-coder:7b-instruct"
      ]
    }
  ],
  "windsurf.activeModel": "deepseek-coder:6.7b-instruct"
}

Step 3: Select the Local Model in Cascade

Open the Cascade chat panel on the right side of your editor. In the model selection dropdown at the bottom of the panel, you will now see "Local Ollama" listed as an option. Select your pulled model, and begin coding!

Performance Benchmarks: Local vs. Cloud

To evaluate the viability of local AI, we ran a series of code generation tests comparing DeepSeek-Coder:6.7b (Local) against Claude 3.5 Sonnet (Cloud) on an Apple M3 Pro (18GB RAM).

*Verdict*: Local models are significantly faster for rapid autocomplete and short edits due to the lack of network roundtrips. However, for massive, multi-file refactoring tasks, cloud models like Claude 3.5 Sonnet remain superior due to their larger context windows and higher reasoning accuracy.

Troubleshooting & Performance Tuning

If you experience performance issues or connection errors, try these optimizations:

• Connection Refused (11434):

  Ensure Ollama is running in the background. On Windows, verify the Ollama icon is visible in your system tray. On macOS, ensure the Ollama application is active.

• Slow Generation (CPU Fallback):

  If tokens are generating at less than 5 tokens/sec, Ollama may have failed to allocate the model to your GPU. Verify that your graphics drivers are up to date and that you have closed other VRAM-heavy applications.

• Context Window Crashes:

  Local models default to an 8k context window. If you send a massive file, the model will fail or forget previous instructions. Keep your active files small and use a .gitignore or .cursorignore to exclude large build directories from indexing.

Conclusion: The Hybrid Developer Setup

Configuring Windsurf Cascade to run local models via Ollama is one of the most empowering setups a developer can build in 2026. It provides a zero-cost, lightning-fast, and completely private environment for your daily coding tasks, while preserving the ability to switch back to cloud models when you need high-level systems architecture reasoning.

By mastering this hybrid workflow, you can optimize your development speed, protect your intellectual property, and eliminate subscription overhead.

To learn more about configuring persistent rules for your local AI agents, check out our guide on CLAUDE.md Configuration, or read our in-depth tutorial on Aider AI Setup to master the terminal.