Ollama runs open-source large language models locally with a simple HTTP API.

On first deploy, a one-shot init job pulls llama3.2 automatically. Models persist under /root/.ollama. Override with OLLAMA_MODEL if needed.