The extension can provide real-time inline code completions powered by a locally running Ollama model as you type.

Enabling Completions

1. Open Settings (Cmd+, / Ctrl+,) and search for ollama.completionModel
2. Enter the name of a locally installed model (e.g., qwen2.5-coder:1.5b)
3. Ensure ollama.enableInlineCompletions is set to true

Completions will appear as grey ghost text in the editor. Press Tab to accept, Escape to dismiss.

Recommended Models

Smaller, faster models work best for completions — they respond quickly enough to be non-intrusive:

Model	Size	Notes
qwen2.5-coder:1.5b	~1 GB	Excellent code quality for its size
qwen2.5-coder:3b	~2 GB	Good balance of speed and quality
deepseek-coder:1.3b	~1 GB	Fast, solid for common patterns
starcoder2:3b	~2 GB	Broad language support

Larger models (7B+) can be used but may introduce noticeable latency between keystrokes.

Fill-in-the-Middle (FIM)

When a model supports FIM (fill-in-the-middle), the extension passes both the text before and after the cursor, giving the model bidirectional context. This produces much better multi-line completions.

All of the recommended models above support FIM.

Disabling Completions

Toggle ollama.enableInlineCompletions to false in Settings at any time. You can also disable it temporarily via the status bar language selector menu.

Troubleshooting

No completions appearing

• Verify ollama.completionModel is set to a model that is installed locally
• Verify Ollama is running (curl http://localhost:11434)
• Check the Ollama output channel (Ctrl+Shift+U then select "Opilot") for errors

Completions are slow

• Switch to a smaller quantized model (1–3 B parameters)
• Verify no other processes are competing for GPU/RAM
• Consider increasing ollama.contextLength to a smaller value to reduce the model's working context