How to install Ollama and start a local AI chat
Original publication: · Sergey Zhadov
This is a translated archive article. The date and pricing discussion refer to the original publication, not a new announcement today.
Ollama runs language models on your computer and provides a terminal chat plus a local API. Install it, choose a model that fits the hardware, launch one model and ask a simple question. Do not download a monster model immediately or expose port 11434 to the internet. That is how a beginner gets a frozen machine or a security hole instead of useful local AI.
Understand what Ollama actually does
Ollama downloads and runs models, stores their files and exposes a terminal chat and HTTP API. It is a runtime and server, not one proprietary model. Results depend on the selected model, size, quantization, prompt and hardware.
Local execution is useful for experiments and data that need not go to a cloud service. It does not guarantee privacy by itself. Plugins, terminals and surrounding applications may still save or transmit information.
Check memory and storage first
Model files take gigabytes, and large variants can consume tens or hundreds of gigabytes. Runtime needs RAM or VRAM. Longer context and parallel requests increase that demand, so leave room for the operating system.
Do not choose only by parameter count. A compact instruct model is a better first test. If the system swaps, fans scream and each answer takes forever, the model is not thinking deeply. The computer is drowning.
Install on Windows, macOS or Linux
On Windows, the installer adds the ollama command and starts the application in the background. Current documentation specifies Windows 10 22H2 or later. macOS needs version 14 or later. Apple M supports CPU and GPU, while Intel Mac runs on CPU.
Linux can use the install script or manual package, then run the server directly or as a systemd service. Do not paste administrator commands blindly. Check the source and script first.
Start the first terminal chat
Open a new terminal and run ollama. The interactive menu can launch a model. A classic run command names the model and opens a conversation. The first start is slower because the model file must download.
Begin with a short verifiable request and one follow-up. Use neutral text or harmless sample code, not company secrets. This is enough to check context and speed.
Test the local API
Ollama serves its API on localhost:11434. The chat endpoint receives a model name and message list. Send one non-streaming test and inspect the status, content and timing fields.
Keep the address and model in configuration, then add a timeout, request-size limit and a clear error when the server is down. Otherwise the interface hangs and everybody wonders what the hell broke.
Do not expose the server carelessly
The default 127.0.0.1 binding limits access to the same computer. OLLAMA_HOST can change that, but another address may expose the service to the network. Configure firewall, authentication, encryption and allowed clients first.
Never forward the port publicly on a whim. An unprotected API can let strangers consume hardware and run installed models. Use a controlled proxy or VPN and retain access logs.
Manage model storage before the disk fills
On Windows, models normally live under the user profile and OLLAMA_MODELS can move them. Paths differ on macOS and Linux. Stop the service and preserve permissions before moving files.
Delete unused models and check free space before pulling another size. Five nearly identical variants are not a strategy. A full system drive will break more than Ollama.
Verdict: one model and closed localhost first
A clean first run is simple: official install, one suitable model, short chat and a local API test. Then measure speed, memory, quality and context before adding interfaces or agents.
The usual mistake is building a home cloud before understanding resources and networking. Keep the service on localhost and match model size to hardware. Then local AI becomes a tool rather than an expensive heater with a blinking terminal.
Translation reviewed:
