Fill-in-the-Middle Playground

The FIM Playground is a small, focused page for testing code-completion-style generation. You provide a prefix (code before the cursor) and a suffix (code after the cursor), and the model fills the gap in between.

Sidebar item: Code Completion (under Tool Workshop) Route: /fimAdded in: v0.4.7 (May 2026)

What is fill-in-the-middle?

Most chat APIs only support left-to-right generation: you give a prompt, the model continues from where you left off. FIM is different — the model sees what comes after the gap as well, and writes content that fits naturally between the two anchors.

This is what powers tools like GitHub Copilot's mid-line suggestions and Cursor's "tab completion." Several Ollama-served models support it natively:

• Qwen2.5-Coder (all sizes)
• CodeLlama (the -code variants)
• Codestral
• DeepSeek-Coder
• StarCoder

Ollama exposes FIM through a suffix field on the /api/generate endpoint. LLMxRay surfaces this in the FIM Playground page.

Using the page

1. Pick a model from the dropdown — coding-capable models are listed first under "Coding models (FIM-capable)".
2. Type your prefix (the code that comes before the cursor) in the left textarea.
3. Type your suffix (the code that comes after the cursor) in the right textarea.
4. Click Generate.

The model streams the missing middle in real time. After generation completes you also see a Stitched preview — the prefix, the generated middle (highlighted in accent color), and the suffix rendered as one continuous block, so you can read the whole result as it would appear in your editor.

Metrics

When generation finishes, four metric cards appear below the output:

• Total duration — wall-clock time from request to the final token
• Model load — how long Ollama spent loading the model into memory (0 ms for warm starts)
• Generated tokens — count of tokens the model produced for the middle
• Speed — tokens per second

If the model doesn't support FIM

If you pick a non-coding model, the page warns you up front. If you try anyway, Ollama returns a clear error: qwen2.5:3b does not support insert (or similar, with the model name in place). The page surfaces this in the output area without crashing.

Behind the scenes

The page calls /api/generate with the new suffix field that Ollama exposes:

{
  "model": "qwen2.5-coder:7b",
  "prompt": "def fibonacci(n):\n    if n <= 1:\n        return n\n    ",
  "suffix": "\n\nprint(fibonacci(10))"
}

The streamed NDJSON response is parsed the same way as a regular chat — one JSON object per line, accumulating response deltas until done: true.

See also

• Chat Diagnostics for general chat
• Tool Workshop for building agentic tools