Ollama's 5-Minute Default Caused 214 Model Reloads in a Single Day

A developer running a local AI chat app discovered that Ollama's default 5-minute idle timeout was silently evicting models from VRAM, forcing a costly cold reload from disk on every subsequent request. Over 24 hours and 1,180 requests, the system logged 214 model load events, with cold starts taking 11.4 seconds to first token compared to just 0.9 seconds when the model was already warm. A cron job set to run every 10 minutes was hitting the 5-minute timeout every single time, meaning it had never once used a warm model. The root cause was partly obscured because passing keep_alive in the request body has no effect on Ollama's OpenAI-compatible endpoint — only the native API honors it. Setting the server-side environment variable OLLAMA_KEEP_ALIVE=24h and moving embeddings to a separate CPU-only instance reduced daily model reloads from 214 down to just 9.