Why I started running LLMs locallyI work with artificial intelligence every day, and for a long time I took for granted that the only way to use language models was through cloud APIs: OpenAI, Anthropic, Google. It works, sure. But it has a cost — and I'm not just talking about money. Every query you send to a cloud service is data leaving your machine. When you work on company code, confidential documents, ideas not yet published, this is a real problem. Then there's latency, rate limiting, dependence on a stable internet connection. Then I discovered Ollama. And I realized the landscape had changed.What Ollama is and why it worksOllama is an open source tool that lets you download and run large language models directly on your computer. The interface is deliberately simple: ollama run llama3.1 — this command downloads the Llama 3.1 model (if not already present) and opens an interactive session. No API key, no account, no data leaving your machine. Behind the scenes, Ollama handles model quantization (which makes models much lighter without significant quality loss), memory allocation between RAM and VRAM, and GPU acceleration when available.The models I use mostLlama 3.1 (8B) is my starting point for general use. It runs comfortably even on machines with 16GB RAM and handles Italian well. Deepseek Coder is what I prefer for code — trained specifically on programming datasets, the quality of technical explanations is noticeably superior. LLAVA is the multimodal model that impressed me most: you can pass it images — error screenshots, diagrams, whiteboard photos — and it analyzes them in natural language.Open WebUI: the interface that makes everything accessibleThe command line is fine for quick testing, but for daily use I want an interface close to ChatGPT. Open WebUI delivers this, deployable with a single Docker command. The key parameter is --add-host=host.docker.internal:host-gateway to connect Open WebUI to the Ollama server running on the host — a detail I discovered the hard way the first time I tried this setup.Modelfile: customizing model behaviorA feature many ignore is the Modelfile, which lets you create customized variants of an existing model with a specific system prompt, temperature, and context window. I have custom models for different use cases: one for technical writing, one for code review, one for generating quiz questions for my courses.Hardware requirements8GB RAM handles 3-4B parameter models well. 16GB RAM is the sweet spot for 7-8B models. 32GB RAM opens up 13-14B models with noticeably better quality. With a dedicated GPU (8GB+ VRAM), everything becomes much faster. Without a GPU, smaller models (7-8B) are still usable daily on modern hardware.Privacy and data sovereigntyThe strongest argument for local AI is privacy. When you run an LLM on hardware you control, your data never leaves. It's not used to train models. It's not logged on remote servers. For companies working with sensitive data, proprietary code or customer information, this is often the primary requirement. With Ollama the answer is almost always yes. Today I use local LLMs for summarizing technical documentation, generating article drafts, code review on fragments containing business logic, and analyzing error screenshots with LLAVA. About 60-70% of my daily AI queries now stay local.