The hype around local large language models has been deafening. People talk about replacing every cloud service with a model running on a laptop. That is nonsense for most tasks.
What actually works is small, focused automation. I run a 7 billion parameter model on a desktop with a consumer GPU. It handles one thing well: rewriting messy meeting notes into clean summaries. It does this without sending data anywhere. That matters for clients who sign NDAs.
I also use it to classify support emails. Three categories. The model gets it right 95 percent of the time. The remaining five percent I handle myself. The cost is zero beyond electricity.
The key is to stop expecting general intelligence from these things. A local LLM is a tool like a calculator. It does one job reliably if you feed it the right input. Trying to make it write code or generate marketing copy from scratch is a waste of time. The outputs are mediocre.
Start with a boring problem. A repetitive task you hate doing. Write a prompt that describes exactly what you want. Test it until the output feels boring. Then automate that one step. Do not expand. Do not add more features.
Local models are good for privacy and latency. They are bad for anything that requires creativity or real reasoning. Use them for what they are. Stop pretending they are something else.
Paul