The first response can be slower because Ollama loads the model into memory. A Reading data indicator or the typing indicator means XLlama is preparing your data or waiting for text; speed depends on your computer.
- Try a smaller downloaded model in Models.
- Test with a few cells and a short question. Large selections and long conversations require more work. In the chip above the message box, choose Selection instead of Active sheet or Workbook.
- Lower Settings → Context size (for example to 8K). Larger sizes let the model read more data but answer more slowly.
- Close other memory-intensive applications.
- In Settings → Keep model loaded, choose a longer period if you use the same model repeatedly.
- Use PROMPT or EXTRACT to reuse cached answers when the inputs match.
Use Stop (or press Escape) to cancel a chat request or Cancel all AI requests for pending work. A cancelled chat that already produced answer text uses one Free response, and the partial answer stays in the chat.
Run Help → Diagnostics to separate a connection problem from a slow model.