Your machine, your data
What 'local' means
When you use ChatGPT, Claude or Gemini, your message travels to the company's servers, a model there writes the reply, and it comes back. A local model is a file of several gigabytes stored on your computer, run by an app such as Ollama, LM Studio or Jan. Nothing is sent anywhere unless you choose a cloud option.
Why people do it
Privacy: confidential documents, health or legal questions, client files. Offline use: on a plane, in a basement office, in places with poor connections. Cost: no subscription once you have suitable hardware. Control: the model doesn't change overnight, isn't withdrawn, and works the same way every time.
The honest trade-offs
Open models you can run at home have improved fast, but on hard reasoning, long documents and the newest knowledge they usually trail the best cloud models. They are slower on modest hardware. Set-up takes an hour, and you handle updates. For many everyday jobs (drafting, rewording, summarising, questions about your own files) they are good enough.
Open weights, not magic
Local models are 'open-weight' models: companies such as Google, Meta, Mistral, Alibaba and DeepSeek publish the trained model file for anyone to download, under licences that vary. Your app loads that file and runs it. Lesson 4 covers choosing one and checking its licence.
What would you ask an AI assistant if you were certain nobody else could ever see the question or the documents?