Completion model is not a code modelThinking and chat models generate prose instead of code in FIM mode. Switch to a Code model (e.g. qwen2.5-coder:7b) for inline completions.
⌨ Best Models for Inline CompletionsFIM · code-only
Only Code models support Fill-in-Middle prompting. Thinking and chat models generate explanations instead of code. Speed estimates at a consumer GPU (8 GB VRAM).