In short: a “smart chatbot” class neural network now runs on an ordinary computer for free and works without the internet. You need a program — LM Studio is the easiest, and it is free for home and for work — plus a model file of a few gigabytes. Download it once, and from then on your processor and graphics card answer your requests, not someone else’s server. Below: five steps in LM Studio, the same path in Ollama, which model to pick for your memory, and where a local model falls short of ChatGPT.
Why keep a neural network at home
A cloud chatbot — ChatGPT, Claude, Gemini — receives every request in full: the text, attached documents, photos. What happens to them next is up to the company and its retention policy. We covered how to use cloud services more carefully in the article on ChatGPT through a VPN, but the service itself sees your conversation either way.
A local model removes that question:
- Requests never leave your computer. The model is a file on disk; the program reads it and computes the answer itself. You need the internet only once, to download the program and the model.
- It works offline — on a plane, off the grid, when a cloud service is down or unavailable in your country.
- Free, with no message limits. You pay only in electricity and computer time.
- You can feed it personal things: contracts, medical records, work code, correspondence — things you would rather not send to a cloud chat.
The honest price: a home model is weaker than flagship cloud ones. It is more than enough for summaries, translation, email drafts, explaining code and going through documents; for complex multi-step tasks it is not always enough.
What your computer needs
The main resource is memory. The whole model is loaded into RAM or into the graphics card’s memory, and its size decides which model you can run. LM Studio’s requirements:
| System | What you need |
|---|---|
| Mac | Apple M1 or newer only, macOS 14 or newer. Intel Macs are not supported |
| Windows | An x64 processor with AVX2 (almost any from the last ten years) or ARM Snapdragon X. 16 GB of RAM and a graphics card with 4 GB of memory are recommended |
| Linux | x64 or ARM64, Ubuntu 20.04 or newer, shipped as an AppImage |
LM Studio recommends 16 GB of RAM; 8 GB works too, but only with small models. On Macs with M-series chips memory is shared between the processor and the graphics, so all of your RAM is the budget for the model, minus what the system itself takes.
How to run a neural network in LM Studio: five steps
- Download LM Studio from the official site lmstudio.ai — it picks the version for your system. A normal install, no sign-up.
- Open Discover — the magnifier icon on the left or
Ctrl+2(⌘+2on a Mac). This is the model catalogue; it searches Hugging Face, the largest store of open neural networks. - Find a model. Type a name such as
qwen3.5 9borgemma 4. You can also paste a direct link to a model page on Hugging Face. Which one to choose — see the table below. - Pick a file variant and press Download. One model comes as many files of different sizes — different degrees of compression (quantisation). LM Studio itself advises taking a 4-bit variant (
Q4in the name) or larger if your memory allows. Next to each variant the program shows whether the model fits fully into your graphics card. - Load the model and type. Go to the chat tab, pick the downloaded model in the list at the top and ask a question. The first answer may take a few seconds while the model loads into memory.
That’s it: you can switch the internet off now, the model answers without it.
LM Studio also has a separate agent app, Bionic, for working with documents and code. It runs both local and cloud models. Cloud models mean someone else’s server and an account, so if privacy is the point, pick models from the local section there.
The same path in Ollama
Ollama is the second popular program for local models. It used to be command-line only; now it has a regular app for Windows and macOS (ollama.com/download). Models come from the Ollama catalogue with one command:
ollama run qwen3.5:9b
On the first run the model downloads, then a chat opens right in the terminal or in the app window. A model from Hugging Face that is not in the Ollama catalogue runs the same way — by its repository address:
ollama run hf.co/<author>/<repository>
By default Ollama takes the 4-bit Q4_K_M variant. Choosing between the two is simple: LM Studio is handier if you want buttons and a visual catalogue, Ollama if other programs on your computer need the model (it has a local API).
Which model to pick
A rule of thumb: the model file should fit into memory with 1–2 GB to spare — the spare goes to the conversation itself (the context) and to the system. If the model does not fit into the graphics card in full, the program splits it between the card and the processor — it will run, but noticeably slower.
| Memory (RAM or graphics card) | What to take | File size |
|---|---|---|
| 8 GB | Ministral 3 3B, Qwen 3.5 4B | 3–4 GB |
| 16 GB | Qwen 3.5 9B, Gemma 4 E4B, Ministral 3 8B or 14B; gpt-oss-20b — at the limit | 6–14 GB |
| 32 GB and more | Qwen 3.8 27B, Gemma 4 26B or 31B | about 16–20 GB |
A little about the families, to find your way around the catalogue:
- Qwen 3.5 (Alibaba, Apache 2.0 licence) — a line from 0.8B up to huge server models; 9B is a good all-rounder for a 16 GB computer. Qwen 3.8 27B came out in August 2026 and is noticeably smarter, but needs about 17 GB.
- Gemma 4 (Google, Apache 2.0) — the lighter E2B and E4B (files of about 5–9 GB) and the large 26B and 31B. With 32 GB of memory, 31B leaves little room for a long conversation.
- gpt-oss-20b (OpenAI, Apache 2.0) — a file of about 14 GB; its developers say it needs 16 GB of memory or more.
- Ministral 3 (Mistral, Apache 2.0) — 3B, 8B and 14B, files from 3 to 9 GB.
The letter B in a name means billions of parameters: the more of them, the smarter the model and the more memory it needs. Ordinary models politely refuse some topics — that is how their developers trained them. There are also models without refusals: the same Qwen or Gemma with that behaviour cut out — what they are and the price of having no refusals is covered separately.
Tainet VPN hides both the contents and the metadata: from the network's side your traffic is indistinguishable from ordinary HTTPS, and the point owner can't see where you go. Sign up through the Telegram bot or in your account on this site, from $0.10 a day.
Try it →If something doesn’t work
The model downloaded but answers very slowly
Most likely it did not fit into the graphics card and the processor does part of the work. Take a smaller model or a more compressed variant (Q4 instead of Q6 or Q8). Close the browser with a dozen tabs and other heavy programs — they use the same memory.
LM Studio won’t start on a Mac
Check the processor: Apple menu → “About This Mac”. LM Studio runs only on Macs with M1 or newer and macOS 14 or newer; it cannot be installed on an Intel Mac. For older Macs, try Ollama.
Answers in Russian are worse than in English
That is a property of the model, not the program. Qwen and Gemma handle Russian well; small 3–4B models are noticeably weaker. If Russian matters, take a bigger model.
What next
Models don’t run only on computers: a modern smartphone can host a neural network on your phone without internet — smaller and simpler, but always in your pocket.
Frequently asked questions
Is it really free?
Yes. LM Studio is free for personal use and for work, Ollama is open source. The models in this article are distributed under free licences. You pay only for cloud models inside these programs, if you choose to use them.
Do I need the internet to use a local model?
Only to download the program and the model. After that everything works offline: your computer handles the requests.
Can the LM Studio developers see my requests?
A local model computes the answer on your computer; the request is not sent anywhere. The exception is cloud models available in some apps (for example in the Bionic agent): their requests go to a server, and the model’s label shows it.
Can I run a neural network without a graphics card?
Yes, on the processor alone, just slower. For a weak computer take 3–4B models — they answer at an acceptable speed even without a graphics card.
How is a local model worse than ChatGPT?
It is smaller: flagship cloud models are trained on more data and run on servers with hundreds of gigabytes of memory. In everyday tasks the difference is small; in complex reasoning and fresh facts it is noticeable — a local model knows the world only as of its training date and does not go online by itself.
A local model keeps what you ask it to itself. Everything else your computer sends to the network — websites, cloud services, downloads — is visible to your provider and the Wi-Fi owner. Tainet hides both the contents and the metadata: sign up in the Telegram bot or in your account on this site, the first 7 days are free, then from $0.10 a day — one subscription for your computer and phone.
