Code

LM Studio: Installing, Configuring, and Launching a Chatbot on Your PC

LM Studio: Installing, Configuring, and Launching a Chatbot on a PC

Learn new skills without spending money: "Practical Course on Neural Networks"

Learn more

LM Studio is cross-platform software that lets you run language models directly on your computer. The application's appearance is similar to ChatGPT: it also features a chat for interacting with the AI ​​and the ability to upload files for analysis. However, the main difference is that all requests are processed locally, so Sam Altman won't be able to see what you're working on.

In this article, we'll cover the installation process for LM Studio, point out where you can download the necessary software, and describe how to run neural networks locally. In addition, we will test one of the models in various scenarios and get acquainted with several alternative services.

Contents

  • Analyzing the characteristics of local models.
  • The LM Studio installation process and subsequent system verification.
  • Beginning to integrate the first neural network.
  • We are in the process of finding the most effective language models for LM Studio.
  • Exploring the capabilities of LM Studio in the context of common everyday tasks.
  • Let's consider various options that can replace LM Studio.

Exploring the characteristics of local models

The key advantage of local language models is confidentiality. All information is stored directly on your device and is not transferred to third-party resources. This is especially relevant when working with sensitive data, such as analyzing corporate financial statements, processing medical records, or drafting legal documents subject to confidentiality obligations.

Other advantages include the following:

  • Adaptation to specific tasks. The local model can be customized to meet individual requirements: it is possible to integrate a knowledge base, download company documents, or teach it specialized terminology. This will allow you to create an assistant that will more sensitively perceive the context of your work and provide more accurate answers.
  • Freedom of communication. All cloud neural networks have certain topics that are subject to discussion restrictions. For example, DeepSeek carefully avoids discussions about Taiwan and the Tiananmen Square incidents of 1989, while ChatGPT often refuses to provide advice on medicine, psychology, and finance. In most local models, such restrictions can be easily lifted.
  • Working without internet access is made possible by local neural networks. They provide assistance in remote regions or situations with an intermittent connection—whether during a flight, at a dacha without Wi-Fi, or in areas with a weak signal.
  • Cost savings. Local models do not require a paid subscription and allow for the ability to process any number of requests without restrictions. Furthermore, everything functions without the lags typical of free cloud services.

A key drawback is the significant dependence on computing resources. For example, working with models containing 70 billion parameters requires at least 64 GB of RAM and a graphics card with a large amount of video memory. Even 32GB of RAM may not be sufficient for local deployment of neural networks of the same caliber as ChatGPT, causing such models to run slowly or not launch at all.

Furthermore, local neural networks do not have internet access. This means that they are unable to find fresh information and, deprived of updates, may invent facts or rely on outdated data.

Read also:

Kandinsky, GigaChat and similar projects: principles of operation of multimodal neural networks

LM Studio installation process and system diagnostics

Go to lmstudio.ai and click the "Download" button. We'll be installing LM Studio on a MacBook Pro equipped with an M1 Pro chip and 16GB of RAM.

If your operating system isn't automatically detected on the website's main page, you should go to the "Download" section. There, select the desired operating system, specify the processor architecture and download the program.

Screenshot: LM Studio / Skillbox Media

After that, open the downloaded file and install it as you would with any other program. On macOS, you need to drag the LM Studio icon to the Applications folder, while Windows users should follow the instructions provided by the installation wizard.

Once the application has successfully launched, you'll see the main screen, which is similar to the interfaces of services like ChatGPT, DeepSeek, Grok, and the like. In the center of the screen, you'll notice a small purple robot labeled LM STUDIO, and below it is an empty field labeled "Send a message to the model…." This is where you'll enter your requests to the model.

Screenshot: LM Studio / Skillbox Media

However, if LM Studio fails to launch, most likely your device does not meet the necessary system requirements:

A few words about system requirements. Recently, we were preparing a material about local neural networks and tested the installation of LM Studio on a MacBook with an Intel processor and on Linux Mint. On the MacBook, the application functioned despite the lack of official support. However, on Linux, the installation went smoothly, but the program failed to launch. The reasons for this remain unclear.

Launching our first neural network

When you first launch LM Studio, you will notice that it is just an interface without integrated neural network models. If you try to enter a query in the chat immediately after installation, you will not receive any results. To start using the application, click the search icon located on the left panel – this will redirect you to the Discover section. Here you will be able to review a list of available open-source models and select the one you like to download. Essentially, it's a kind of marketplace for neural networks.

Discover section in LM Studio Screenshot: LM Studio / Skillbox Media

As an example, let's consider Qwen3 4B Thinking 2507 - a small language model capable of reasoning, developed by the Qwen team. This model can handle contexts up to 256,000 tokens and is ideal for a variety of everyday tasks, such as writing and editing text, programming, explaining solutions, and answering questions. It requires 4 GB of RAM and 3 GB of free hard disk space. Search for Qwen3-4b-thinking-2507. If you're using macOS, choose the MLX format from the available versions, and for other operating systems, choose GGUF. After that, click the Download button, wait for the download to complete, and then select the Use in New Chat option.

Installing a local neural network for macOS in MLX format Screenshot: LM Studio / Skillbox Media

Then a chat window will appear in which you will be able to enter your request for the neural network. We've implemented a reasoning model, and before generating a response, it will briefly analyze your question and select the most appropriate option. The chat will also display the reasoning sequence, allowing you to see why the neural network reached its conclusion or where it might have made mistakes. Please note that the speed of responses depends on the capabilities of your computer. For example, on a Windows system, in addition to sufficient RAM, it is advisable to have a powerful graphics card. This is because LM Studio will use it to process some of the model's layers, which helps speed up generation. If there is no video card or its memory is insufficient, the entire load will go to the processor, which will lead to a slowdown.

Example of the response of the local model Qwen3 4B Thinking 2507 in LM Studio Screenshot: LM Studio / Skillbox Media

In addition to the search function, on the left sidebar There are icons for the folder and the terminal. Clicking the folder icon takes you to the My Models section, which displays all the language models stored on your device. When selecting any model, you can view the developer's name, number of parameters, quantization type, and file size. If any model is no longer relevant, you can easily remove it from this section.

My Model section in LM Studio Screenshot: LM Studio / Skillbox Media

The terminal icon opens the developer menu, where you can view the parameters of the active model, view logs and start the server to integrate the neural network with other applications. This section also provides model settings: you can define a system prompt, adjust the inference process, and change the context length.

Developer section in LM Studio Screenshot: LM Studio / Skillbox Media

Finding ideal language models for LM Studio

In LM Studio lets you work with a variety of language models. Below are some popular options suitable for various situations:

  • gpt-oss-20b is an open-source model developed by OpenAI with 20 billion parameters. It will run on standard PCs with 16 GB of RAM and is ideal for users who value a response style similar to that offered by ChatGPT.
  • gpt-oss-120b is an advanced model with 120 billion parameters that can efficiently solve complex problems, reason, and analyze large amounts of data. It requires a graphics card with 80 GB of video memory to run fully.
  • qwen2.5-coder-14b is a model designed for code generation and technical documentation. It can run on GPUs with 8 to 12 GB of video memory and on systems with 16 GB of RAM.
  • Gemma-3-12b has the ability to generate text and code, as well as recognize images and answer questions. The model, which has 12 billion parameters, requires about 32 GB of RAM. Other versions are available: gemma-3-1b requires 4 GB, gemma-3-4b requires 8 GB, and gemma-3-27b requires 64 GB of RAM. If you decide to use the compressed versions, the latter model can be activated with 16 GB of RAM.
  • phi-4-mini-reasoning is a compact reasoning model developed by Microsoft with 3.8 billion parameters and a larger context window. This model effectively solves problems related to logical inference and planning. It requires at least 16 GB of RAM to run.

The most convenient way to search for local models is to use the Hugging Face portal, rather than the built-in search in LM Studio. This resource can be compared to GitHub for neural networks, as it hosts thousands of open-source models. Each repository contains the neural network itself, information about its functionality, as well as recommendations for launching, retraining, and adapting to various tasks.

Our website now offers filters that make it easier to search for models. You can use them to specify the number of parameters, task types, supported languages, as well as adaptation for specific applications and optimization for provider requests. Then you just go to LM Studio and use the search function to find the desired neural network.

Filter window and list of models on Hugging Face Screenshot: Hugging Face / Skillbox Media

Evaluation of LM Studio capabilities in everyday practice

The effectiveness of local models' responses is significantly The accuracy of the neural network is determined by the number of parameters: the greater their number, the higher the accuracy and quality of the results. In the following examples, we will test the Qwen 3 neural network, which has 4 billion parameters. We will evaluate its capabilities in areas such as programming, information retrieval, text writing, and mathematical problem solving.

At the beginning of our experiment, we asked the neural network to develop a Python script that would display the English alphabet as Aa, Bb, Cc, Dd, and so on. As a result, we received fully functional code that successfully completes the task:

Screenshot: LM Studio / Skillbox Media

We increased the difficulty level of the task and asked Qwen to fix the error in the code. The neural network analyzed the problem, suggested corrected code, and explained why the function wasn't working properly. Here's the text of our request:

When a function returns a new list with an added element, you may be encountering an issue where elements accumulate across multiple calls. This can occur because you're using a mutable object, such as a list, and adding elements to the same object with each function call.

To fix this issue, ensure you're creating a new list with each function call, rather than modifying an existing one. For example, instead of appending elements to an existing list, create a new list that includes all the elements from the old list plus the new element. This will prevent value accumulation.

Here is an example of the corrected function:

«`python
def add_element(original_list, element):
new_list = original_list.copy() # Create a copy of the original list
new_list.append(element) # Add a new element to the copy
return new_list # Return a new list
«`

This way, each time the function is called, a new list will be returned with only the specified element added, without accumulating previous values.

Here is the response from the localization:

The main problem is that the list items=[] is used as a default value. This results in the same list being modified each time the function is called, instead of a new one being created. As a result, repeated calls to the function continue to append elements to the list.

Corrected code:

Explanation:

  • The first time append_item is called with the "a" argument, a new empty list [] is created, then the element "a" is added to it, and the list ["a"] is returned.
  • The second time append_item is called with the "b" argument, the value of the items parameter is checked, which is None by default. In this case, a new empty list, denoted as [], is initialized, and the element "b" is added to it. As a result, the function returns a list containing only "b": ["b"].
  • The third time the append_item function is called with the argument "c", an empty list [] is again created, into which the element "c" is then added.

As a result, with each new call, the function creates and returns a fresh list, instead of appending elements to an existing one.

Example of work:

Our Qwen 3 does not have the ability to search the Internet, which means that it can create fictitious events and distort facts. When we asked what colors Apple's iPhone 17 Pro Max would come in, the neural network responded that it was a model that hadn't been released yet, and therefore had no design information.

Screenshot: LM Studio / Skillbox Media

ChatGPT easily handles this request, referring to to Apple's official website. We can't deny that searching for information online plays a significant role.

Screenshot: LM Studio / Skillbox Media

As part of testing, we made a request to Qwen 3 with a request to create a birthday greeting for a fictional character named Artyom. To make the task more interesting, we specified that the greeting should include original references to Formula 1 and programming—our birthday boy's hobby. Below is the resulting text, which we can pass on to any Artem:

Screenshot: LM Studio / Skillbox Media

Let's consider a complex mathematical expression and ask the neural network to simplify it.

Unfortunately, Qwen 3 does not have an image recognition function. Therefore, we converted the entry to LaTeX format and presented it as follows: \sqrt{(2\sqrt{11}-3\sqrt{5})^{2}}+\frac{2}{4\sqrt{5}+9}+\sqrt{44}-18.

The language model completed the task in 36 seconds, providing a detailed explanation of the solution and achieving the correct result. Here's what she got:

Screenshot: LM Studio / Skillbox Media

The main problem is that the items=[] list is used as a default value. This leads to the fact that each time the function is called, the same list is modified, instead of a new one is created. As a result, with repeated calls to the function, the elements begin to accumulate.

Corrected code:

Explanation:

  • The first time the append_item function is called with the argument "a", a new empty list [] is first formed. The element "a" is then added to this list, resulting in the list ["a"].
  • The second time the append_item method is called with the argument "b", the value of the items parameter is checked, which is None by default. This creates a new empty list [] that has the element "b" appended to it, returning a list containing only that element: ["b"].
  • The third call to append_item with the argument "c" again creates an empty list [] that then has the element "c" appended to it.

As a result, the function now creates a new list each time it is called, instead of accumulating elements.

Example in action:

Exploring Other LM Studio Options

In this section, we aim to provide a brief overview of a number of free, cross-platform applications that offer similar functionality to LM Studio. Based on our analysis, we have identified applications such as Ollama, GPT4All, and Jan.

We installed these applications and intended to load the Qwen3 4B Thinking 2507 model into each of them. We would then ask each of them the same question: "What advantages do you have over LM Studio?" — and present the results.

However, our attempts were unsuccessful. Installing Qwen 3 in GPT4All and Jan was successful, but both programs returned an error when attempting to retrieve a response. The Ollama catalog didn't contain the required model, so we decided to download the base gemma3:4b, which froze during the first launch.

We are not drawing any definitive conclusions from this situation. Perhaps it would be worth examining the settings of each service in more detail, trying restarting them, or testing other models. If you feel inclined, by all means experiment. However, if you prefer to get started without unnecessary difficulties, as of the end of 2025, we can only recommend LM Studio.

On the left is Jan, on the top right is GPT4All, on the bottom right is Ollam. In theory, they are as easy to use as LM Studio. But, apparently, not today. Screenshot: Jan / GPT4All / Ollama / Skillbox Media

Learn more fascinating facts about coding in our Telegram channel. Join us!

Read also:

  • Ten best local neural networks for personal computers and laptops in 2025.
  • Avatar creation: 8 best neural network solutions for image stylization.
  • Yandex Neuro: AI search capabilities and ways to use it.