JAW's Wiki

Install ollama and do some tests from the command line

Get the current version (and proof of life)

ollama -- version

Get a model, this one is a 8b parameter coder model (but always check for better models, it is constantly changing):

ollama pull qwen3-coder:8b

Show all the models currently installed:

ollama list

Now make sure that the API running on the local server is up:

curl http://localhost:11434/api/tags

This setup creates a local coding assistant environment similar in workflow to tools such as Claude Code.

Components:

• Ollama runs local coding LLMs.
• OpenCode provides an agent-style coding interface.
• Qwen 2.5 Coder provides the coding intelligence.
• Git provides version tracking and a safe way for the AI to propose changes.
• Projects can be analysed, reviewed, refactored, and modified locally without sending source code to the cloud.

The target workflow:

1. Open a code project in OpenCode.
2. Ask the AI to analyse, review, or modify code.
3. Review proposed changes.
4. Commit changes through Git.
5. Allow larger models to perform deeper analysis overnight.

The system is CPU-focused. The GPU is too small to significantly accelerate larger models.

Install Ollama for Windows.

Verify installation:

ollama --version

Check installed models:

ollama list

Ollama runs a local server:

http://localhost:11434

Verify the API:

curl http://localhost:11434/api/tags

This should return installed models.

Install:

ollama pull qwen2.5-coder:7b

Good for quick tasks.

Approximate memory usage:

~5GB

Install:

ollama pull qwen2.5-coder:14b

This became the preferred interactive model.

Approximate memory usage:

~10GB

Good balance between:

• reasoning ability
• speed
• memory usage

Install:

ollama pull qwen2.5-coder:32b

Approximate memory usage:

~24GB loaded

Excellent for:

• architecture reviews
• large code analysis
• overnight tasks

However, on a 32GB laptop it leaves Windows with little breathing room.

Ollama context can be increased using:

setx OLLAMA_NUM_CTX 16384

Restart applications after changing environment variables.

Verify:

echo %OLLAMA_NUM_CTX%

Check the active model:

ollama ps

Example:

NAME                 CONTEXT
qwen2.5-coder:14b    16384

For a memory-limited machine, reduce if necessary:

setx OLLAMA_NUM_CTX 8192

Install the Windows OpenCode desktop beta.

OpenCode provides the agent workflow:

• project browsing
• multi-file analysis
• change planning
• editing
• Git-aware workflows

Ollama is the “brain”.

OpenCode is the “agent”.

OpenCode decides:

• which files to inspect
• what prompts to send
• what changes to make
• how to sequence tasks

The LLM itself does not independently browse the project.

OpenCode configuration is stored at:

%USERPROFILE%\.config\opencode\opencode.jsonc

Example configuration:

{
  "$schema": "https://opencode.ai/config.json",
 
  "provider": {
    "ollama": {
      "name": "ollama",
      "npm": "@ai-sdk/openai-compatible",
 
      "options": {
        "baseURL": "http://localhost:11434/v1"
      },
 
      "models": {
        "qwen2.5-coder:7b": {
          "name": "qwen2.5-coder:7b"
        },
 
        "qwen2.5-coder:14b": {
          "name": "qwen2.5-coder:14b"
        },
 
        "qwen2.5-coder:32b": {
          "name": "qwen2.5-coder:32b"
        }
      }
    }
  }
}

The `/v1` suffix is important because OpenCode uses the OpenAI-compatible API.

Git is strongly recommended.

Example:

cd MyProject
 
git init
 
git add .
 
git commit -m "Initial project snapshot"

This gives a clean rollback point before allowing AI changes.

OpenCode can then:

• inspect the repository
• create plans
• modify files
• show diffs

A small intentionally-buggy C# project was created to test the workflow.

The AI was asked:

Do not modify files.

Analyse the entire project structure.

Create REVIEW.md containing:

1. Project overview
2. Architecture summary
3. Bugs found
4. Potential runtime issues
5. Security concerns
6. Code quality issues
7. Recommended order of fixes

The model successfully:

• mapped the project
• identified bugs
• produced a structured review
• created a refactoring plan

Best everyday model.

Advantages:

• fast enough
• good reasoning
• comfortable memory usage
• suitable for interactive coding

Advantages:

• deeper analysis
• better architectural reasoning
• useful for large code reviews

Disadvantages:

• consumes ~24GB RAM
• leaves Windows with ~1.7GB free
• can cause paging and sluggishness

Best use:

1. Start a task.
2. Let it run.
3. Review results later.

Symptoms:

404 Not Found

Solution:

Check Ollama first:

ollama list

Confirm exact model name:

qwen2.5-coder:32b

The model ID must exactly match.

Check the API:

curl http://localhost:11434/api/tags

If models appear, Ollama is working.

The OpenCode provider should use:

http://localhost:11434/v1

not:

http://localhost:11434

After:

setx OLLAMA_NUM_CTX 16384

This:

echo %OLLAMA_NUM_CTX%

will not update in the same command window.

Open a new terminal.

Check memory:

Get-Counter '\Memory\Available MBytes'

Example:

Available: 1700 MB

This means the machine is under pressure.

Unload:

ollama stop qwen2.5-coder:32b

Use 14B for normal work.

Configure Git:

git config --global user.name "Your Name"
 
git config --global user.email "you@example.com"

Then retry:

git commit -m "Initial project"

The real configuration is easier to edit manually:

%USERPROFILE%\.config\opencode\opencode.jsonc

For this laptop:

Context:

16384

Workflow:

Git snapshot
    ↓
OpenCode analysis
    ↓
AI review/plan
    ↓
AI changes
    ↓
Review diff
    ↓
Commit

This provides a capable local coding assistant while keeping source code entirely on the local machine.