Local LLM Guide: Run AI on Your PC with LM Studio

A local LLM is an AI model that runs on your own PC instead of a provider's servers. Its biggest advantages are that what you look up never leaves your machine and that you can keep using it without a monthly fee.

A 7B-class model, the usual starting point, runs on a PC with about 8 to 16GB of memory. The first setup takes roughly 30 to 60 minutes, and after that you can chat without an internet connection. This guide covers how local LLMs differ from hosted AI services, what each costs, how much PC power you need, how to choose a model, and how to start with the free app LM Studio.

---

Should you choose a hosted AI service or a local LLM?#

If you are unsure, start with a hosted service. The real deciding question is whether you want your input kept off external servers. If that matters most, choose local. If you mainly want to try AI with no effort, choose a service.

There are two options:

• Hosted service (such as ChatGPT): open it in a browser, create an account, and you can start the same day. Prices run roughly 1,000 to 3,000 yen per month.
• Local LLM: the model runs inside your own PC. If your current PC can run it, the extra cost is close to zero and there is no monthly fee, but you need enough PC power and some effort for the first setup.

Local also lets you pick which model to install, and developers around the world publish models with more freedom in how they respond.

What is a local LLM, and how does it keep your input private?#

A local LLM runs entirely inside your PC, so your conversations never reach an external server. LLM stands for large language model, the core of an AI that has learned from large amounts of text and can answer questions and write documents.

The difference resembles browser games versus packaged games. A browser game lives on the operator's server, needs a connection every time you play, and stops working when the line drops. A packaged game is installed on your own machine, runs without a network once installed, and does not disappear when the service shuts down. A local LLM is the packaged kind: you download the model file, put it on your PC, and keep the whole conversation there.

The main benefit is that what you look up stays on your PC. Even the record of what you asked is never left on an outside server. Many hosted services say they will not use your input as training data, which is an important promise, but your text still passes through the operator's infrastructure. "We won't use it" and "it never reaches us" are different things. With a local LLM, once the model has downloaded, chatting needs no communication at all.

This matters in cases like these:

• Confidential work: summarizing materials a client entrusted to you or an unreleased plan. Some workplaces forbid entering such material into external services by contract.
• Personal research: hobby questions you would not raise with family or friends, or delicate source-gathering for creative work, explored at your own pace where no one else sees it.
• Searches you would rather not have on record: health worries, money problems, relationship trouble. The questions you hesitate to type into a search engine are often the ones you most want to talk through with an AI.

Hosted services still win on convenience and response speed.

What are the hurdles of running a local LLM?#

There are three hurdles: PC power, choosing a model, and the setup effort. None is unusually hard, but each is easier to handle if you know it in advance.

How much GPU and memory does your PC need?#

Your PC's power directly decides how large a model you can use and how fast it responds, because all the AI's calculation happens on your machine. Two parts matter most:

• GPU: a processing unit for graphics. The more powerful it is, as in a gaming PC, the faster the AI calculates.
• Memory: the equivalent of the size of your work desk. The model cannot run unless it fits here.

"B" stands for billion, the number of values inside the model. More values make the model smarter, but heavier to run.

How do you choose a model's size and lineage?#

Choose along two axes: size and lineage. A bigger model gives better answers but demands more from your PC, so beginners should start with the largest size their own PC can run.

Lineage matters because several build styles exist for models that do not cut off your research. Three are representative:

• Dolphin family: the "education" style, in which the model is trained from the start with additional training. It is the original.
• huihui family: the "workshop" style, which modifies the weights (the numbers etched in by training) of a finished model afterward. When a new model becomes a topic, a modified version tends to appear within days to weeks.
• OBLITERATUS family: the "surgery" style, which observes the model's internals, studies the structure of refusal, and then removes it (OBLITERATUS toolkit). For the technique behind this, see what abliteration is.

Even with the same base model, a different modifier and a different quantization make it effectively a different model, so check the author name and tags after the model name.

Quantization deliberately coarsens the precision of the weights to shrink the file, and the lightened version is called a quantized version. With a label such as "Q4," you accept some loss of quality in exchange for a model that runs lighter. In LM Studio you narrow your search by this quantization level, and our GGUF quantization guide explains the formats.

How much effort does the software setup take?#

The two standard tools are LM Studio and Ollama. Both are free and handle searching, downloading and running models. The first setup takes 30 minutes to an hour, and the steps are not hard, but you do need to be comfortable fiddling with your PC.

Which suits you better, local or hosted?#

If your current PC can run it, local gets cheaper the longer you use it, while a hosted service is easier if you are still testing the waters. Your usage frequency is the quickest way to decide.

Who should choose a local LLM?#

• People whose top priority is keeping what they look up off external servers
• People who enjoy fiddling with a PC
• People who plan to use AI heavily over a long period

Who should choose a hosted service?#

• People who want to start right now
• People unsure about their PC's specs
• People who want to try it casually first

How do the costs compare?#

• Local (using your current PC): almost no extra cost, only a slightly higher electricity bill.
• Local (buying a new PC): over 100,000 yen. A desktop with a GPU is the realistic option.
• Hosted service: roughly 1,000 to 3,000 yen per month.

At 3,000 yen a month, a year costs 36,000 yen. If your current PC can run the model, local is cheaper, and if you would need a new PC, over 100,000 yen equals several years of service fees.

What if you are still unsure?#

Start with a service, and move to local if it starts to feel limiting. There are two reasons:

• You cannot know your usage until you try. Some people use AI every day, while others need it a few times a week. Checking your habits on a service first avoids wasted spending.
• Apart from "your input stays off external servers," a hosted service is stronger. For the latest model performance, response speed and range of features, a service is the more comfortable way to begin.

How do you start running a local LLM with LM Studio?#

You can start in four steps, taking 30 minutes to an hour the first time. After that, adding a model takes only a few minutes.

Step 1: How do you check your PC's specs?#

First find out how powerful your PC is by checking whether it has a GPU and how much memory it has. On Windows, use Task Manager, and on a Mac, use "About This Mac". Compare the result with the table above to get a sense of which sizes it can probably run.

Step 2: Which software should you install?#

There are two standard choices, and both are free:

• LM Studio: the standard for beginners, with a screen you operate by mouse (a GUI). Searching, downloading and chatting all happen in this one app, and if you are unsure, it is enough to start with.
• Ollama: a type you operate by typed commands. Its screen is plain but it is lightweight, and some people find it faster once they are used to it.

Download either from its official site and follow the on-screen instructions.

Step 3: How do you download a model?#

In LM Studio, type a model name into the search box and candidates appear. Narrow them by quantization level, choosing a light version such as Q4 that suits your PC.

If you want a model with more freedom in how it responds, search for names from the Dolphin, huihui or OBLITERATUS families, and see our overview of uncensored models. At first, one model sized to run on your PC is enough.

Step 4: How do you run it?#

Load the downloaded model and type a question into the chat box. When a reply comes back, you have succeeded. Start with simple questions to check the response speed and quality.

---

Where do beginners get stuck with local LLMs?#

Beginners get stuck in three places, and none is a fault or a setup mistake. Each has a simple fix.

1. A large model loads and will not run. Someone installs a 32B model because they want something powerful, and it stalls while loading. The cause is simply that the PC lacks the power the model demands. Start with a 7B-class model, which runs on a PC with 8 to 16GB of memory, and move up gradually if there seems to be room.

2. The download does not finish. Model files range from a few GB to dozens of GB. On an ordinary home connection, one file can take tens of minutes to several hours, so it is best to start downloads when you have time, such as overnight.

3. Responses appear slowly. Text that shows up a little at a time can look like a breakdown, but it is normal in most cases. Your PC is working flat out, and response speed is proportional to its power, which shows up most on laptops. If the slowness bothers you, switch to a version with lighter quantization or a model one size smaller.

What the three have in common is that the software is not hard to operate, but you have to search for a fit between your PC's power and the model's demands. Whether you find that search fun or tedious is what divides local from hosted. If you enjoy trying different sizes and quantizations, local suits you, and if the trial and error wears you out, you can go back to a service without forcing yourself.

Which tool should you run a local LLM with, and can you use the models commercially?#

Besides LM Studio and Ollama, TechLabs (2026-05-10) and UNICORNEE (2026-06-16) also list Jan, llama.cpp and vLLM. They differ mainly in interface and intended use, and all of them can serve an OpenAI-compatible local API.

Licenses decide whether you can use a model commercially. According to TechLabs (2026-05-10) and UNICORNEE (2026-06-16), Qwen and Mistral use Apache 2.0, which is commercial-friendly, and Phi and DeepSeek use MIT, which is permissive. Meta's Llama community license sets conditions, including limits on large-scale services, and Google's Gemma has its own terms. Check each model card before commercial use.

What is Deepent Research?#

Deepent Research is a browser-accessible LLM chat service built to refuse as little as possible. It keeps going with research that a standard AI would cut off mid-answer.

• It refuses nothing. We run uncensored and abliterated models with guardrails reduced as far as possible, so prompts a standard AI answers with "I can't help with that" still get a response (Qwen3.5 / Qwen3.8-27B-abliterated, depending on your plan).
• Your input is never used for training. What you look up is used only to provide the service, never to train AI or for any other purpose. Personal data is protected, and our privacy policy is public.
• Try it free. No GPU and no install, just open a browser. In the logged-out trial your conversation isn't saved, so you can start casually.

→ Try it free, no sign-up: https://deepentresearch.com

No question can stop research.

Frequently asked questions#

Are local LLMs free to use?#

⌄

If your current PC can run the model, the extra cost is close to zero and there is no monthly fee. Your electricity bill rises slightly. If you need to buy a new PC, a desktop with a GPU costs over 100,000 yen. A hosted service, by comparison, costs roughly 1,000 to 3,000 yen per month.

What PC specs do you need to run a local LLM?#

⌄

A 7B-class model, the usual starting point for beginners, runs on a gaming-class PC with about 8 to 16GB of memory. A 13 to 14B model needs more than 16GB, and a 32B model needs a large amount of GPU memory, meaning a serious desktop GPU. Larger models answer better but demand more from your PC.

Can you use a local LLM without an internet connection?#

⌄

Yes. Once the model has finished downloading, chatting requires no network communication, so a local LLM works even where there is no connection. What you look up stays on your PC and never reaches an external server, which is the main advantage over a hosted service.

Is setting up a local LLM difficult?#

⌄

With a free app such as LM Studio, the steps themselves are not hard. The first setup takes about 30 minutes to an hour, and adding a model later takes only a few minutes. You do need to be comfortable fiddling with your PC, though, because you have to match the model's size to your hardware.

Disclaimer#

This article was drafted with AI assistance and then reviewed, fact-checked, and edited by the Deepent Research team before publication.
It is provided for educational and research purposes and is not intended to encourage unlawful activity or crime.
Where figures are cited, they come from third-party public reports and vendor-published values and may vary with measurement conditions.