Transcription
When we use neural networks, we must remember, everything goes to the servers of a foreign company. We have already shown tools that help run models locally. Today, about another one. Ollama. You download it once and you have your own personal assistant on your computer. Without a subscription, without the internet, and confidentially. True, with nuances regarding hardware requirements. Ollama appeared in 2023 as a purely terminal tool. But in July 2025, a version with a full-fledged desktop application for MacOS and Windows was released. Now it's no longer a terminal, but a familiar interface. A chat window, a list of models, a history of dialogues. Approximately like any other chatbot, only everything works locally on your hardware. The essence of Ollama is simple. You download the model once and then use it without restrictions. Installation takes literally a few minutes and will not cause difficulties. Go to the website, click Download. The site itself detects your operating system. Ollama works on MacOS, Windows, and Linux. Launch the installer, follow the instructions, and that's it. After installation, Ollama lives in the tray and runs in the background. You can open it through the application. The interface is minimalistic. On the left, a button for a new chat and conversation history. In the center, the familiar dialogue window and a drop-down list with available models. Among the key settings, it is worth paying attention to automatic model updates. Useful, but they are large, this needs to be kept in mind. Context length, roughly speaking, is how much text the model keeps in its head for one conversation. The longer it is, the more detailed the answers in long dialogues, but the more memory is consumed. This depends on your hardware. The ability to use cloud models. To do this, you will need to log in on the Ollama website. Briefly about the difference. Cloud model. Your request flies to the company's servers, is processed there, and returned to you. The ability to try faster, full-fledged models. But we forget about privacy. Local models are downloaded to your disk and run entirely on your processor and graphics card. No leaks for free, but with a caveat about hardware. Ollama supports both options, but its main value is precisely local models. If privacy is important to you or you want to work without the internet, this is your choice. How to install a model? Firstly, you can simply choose from the suggested ones, and when you access it, it will download itself. But this is not the entire list. The model library is on the Ollama website. There are many ready-made options. If you click on View All, you can see all versions. MLX is a mark that the model is optimized for Mac. Models differ in parameters and sizes, respectively. The larger it is, the smarter the model, but the more memory it requires. Here you will need powerful hardware. Focus on the parameters of your graphics card and RAM, but the rules of the golden mean are still in effect. Quantization. Another parameter is model compression. The original model weighs a lot, but the quantized version in 4 bits Q4 takes up several times less, albeit with a loss in quality. By default, Ollama uses four-bit quantization. You can install a model in two ways. Commands are written in the terminal if you go to the model page, or directly from the application. Simply copy the name from the website and go to the application. Paste it into the drop-down list, find it, and select it. The first time you access it, the download will begin. The model is downloaded once. After that, it lives on your disk and works without an internet connection. Let's test some text query first. This is the detailed answer I received. And depending on the model, there are other possibilities. In particular, to attach some file. It will understand the city from the picture. Yes, it already understood something. Now let's try testing a cloud model. I'll repeat again, here all requests go to the developers' servers. To connect, you will have to register on the site, and by phone number. Only after that can you work fully. We send the request and get the result. Now about a feature that appeared in January 2026 and immediately became a topic of conversation among developers. The Ollama Launch command. This is launching AI assistants directly within your code. What tools are supported? As of today, these are Claude, OpenCode, Codex, and Droid. Each of them is a separate application that you need to install yourself. With one command, Ollama configures and launches them and connects them to your local or cloud model. Ollama is like simply downloading a program to your computer. Install it, choose a model for your hardware, and use it. No accounts, no payments, no limits. Yes, it won't replace neural networks for complex tasks. The power is still limited by your computer. But for most everyday things, writing text, understanding a topic, working with sensitive information, or simply testing models, why not? That's all. Perhaps, in the modern world, the most important thing is to remain human in all senses. Alexander Rybakov, Teplitsa of Social Technologies.