Gemma 4 Structured-Task Performance: Field Report from a Local-First App

Benchmark data and prompt-format findings from deploying Gemma 4 E4B in a real application. Intended for LLM teams (Gemma, Ollama) and developers building structured-output pipelines on local models.

Context

We build Gary, a privacy-first personal assistant CLI for macOS. It runs entirely locally — an encrypted database, a daemon process, and a local LLM via ollama. The LLM handles three structured tasks: