Cloud models are strongest for broad reasoning. Local models win on privacy, offline flights, and high-volume experiments you do not want billed per token.
Good local use cases
- Summarising long internal docs
- Draft commit messages and changelog entries
- Regex/log triage on sensitive client data
- Prototyping prompts before sending to GPT-4 class models
Poor local use cases
- Large multi-repo refactors without retrieval
- Legal/tax advice (YMYL)
- Autonomous production deploys
Practical Mac setup
- Install Ollama
- Pull a mid-size coder model (match RAM — 16GB vs 32GB matters)
- Pair with IDE only for bounded tasks
- Keep verification scripts unchanged