Running Qwen3 on Windows with LM Studio

Who this is for and what you will be able to do

This is for readers on Windows who prefer a graphical interface and want to run Qwen3-8B on their own computer. After working through it, you will download and load the model in LM Studio and chat with it, install the model manually if your network is slow, and turn on LM Studio's local server to call the model from a few lines of Python.

What LM Studio is and how it differs from Ollama

LM Studio is a desktop application for running large language models locally, with a graphical interface. The guide describes its interface as richer than Ollama's, with a strong model ecosystem, and suited to use without a network connection^[1]. Ollama is mainly managed from the command line, while LM Studio handles searching, downloading, chatting and running a server within one window.

Preparation: your computer and the Python environment

LM Studio recommends models based on your hardware. The guide's author used an RTX 4070 Ti and could run every size of the model^[1]. With 8 GB of VRAM, the guide recommends the smaller Q4 or Q3 files, as covered below.

To call the local API you need Python greater than 3.7 and the openai library, installed with pip install Openai^[1]:

pip install Openai

The openai library here only talks to the local server. You do not need an OpenAI account for this.

Step 1: install LM Studio and download the model

Go to lmstudio.ai and click download. Installation is a matter of clicking through the steps^[1].

Once it is open, click the magnifier icon at the lower left, search for Qwen3-8B, and choose the version marked with a green rocket. LM Studio marks each model by hardware fit: a green rocket means it runs smoothly, no rocket means it runs slowly, and a red mark means it cannot run at all^[1]. Then click the download button at the lower right^[1].

Step 2: install the model offline when the network is slow

If downloads are slow, you can use the GGUF files from ModelScope and put them into LM Studio's model folder by hand.

Find the model storage path in LM Studio first, then create a folder named Qwen3-8B-GGUF under the lmstudio-community directory^[1].

Next, open the unsloth/Qwen3-8B-GGUF repository on ModelScope at https://www.modelscope.cn/models/unsloth/Qwen3-8B-GGUF/files^[1]. Which file to pick depends on your VRAM:

• With more VRAM, the guide's author chose the Q8 quantized file Qwen3-8B-Q8_0.gguf on an RTX 4070 Ti^[1].
• With 8 GB of VRAM, the guide recommends the Q4 file Qwen3-8B-Q4_K_M.gguf or the Q3 file Qwen3-8B-Q3_K_M.gguf^[1].

After the download, put the file into the Qwen3-8B-GGUF folder. Back in LM Studio, the model appears in the list^[1].

Step 3: chat with Qwen3 in the interface

Click the yellow chat tab, then click "Select a model to load" and choose Qwen3-8b from the dropdown^[1]. The first load takes a while while the model is read into memory.

Step 4: turn on the local server and call it from Python

Open the green terminal module in LM Studio, click Settings, and turn on "Serve on Local Network"^[1].

Then select and load the model. The panel on the right shows the model name and the server address, which you need for the call^[1].

The call code is below. base_url points at the local server, api_key can be any string, and model takes the model name shown on the right:

from openai import OpenAI

# Point to the local server
client = OpenAI(base_url="http://localhost:1234/v1", api_key="lm-studio")

completion = client.chat.completions.create(
  model="qwen3-8b",
  messages=[
    {"role": "system", "content": "你是一个回答我日常问题的ai辅助机器人"},
    {"role": "user", "content": "如何做宫保鸡丁。"}
  ],
  temperature=0.7,
)

print(completion.choices[0].message.content)

In this code, qwen3-8b is the model name, and localhost:1234 works when you call from the same machine. For a computer on your local network, replace the address with the LAN IP shown in LM Studio^[1].

The example sets a system message that makes the model a daily-question assistant, asks how to make Kung Pao chicken, and uses temperature 0.7^[1].

The sample response starts with a <think> block that holds the model's reasoning, followed by the final answer^[1].

Common questions

Qwen3-8B does not show up in the model list. Check the search term, and check whether the marker is red. A red mark means the computer cannot run it, so choose a smaller quantized version^[1].

The model file you copied in does not appear in LM Studio. Check that the file is inside the Qwen3-8B-GGUF folder and that it is a complete .gguf file, then look again in LM Studio^[1].

Another computer on the LAN cannot connect. Make sure "Serve on Local Network" is on, and use the LAN IP shown in LM Studio rather than localhost^[1].

**The answer begins with a long <think> block.** That is the output format of thinking mode, and the answer comes after it^[1].

Limits

The guide uses Qwen3-8B as its example, and an 8 GB card needs a smaller quantized file. LM Studio's markers are a hardware-based guide, and real speed depends on the machine. For offline installs, file names and versions on ModelScope can change, so check the actual files on the page before downloading. The code here shows one chat request and includes no error handling.

Next steps

For more LM Studio settings, see its official site at lmstudio.ai. If you prefer the command line, compare it with this site's entry on Ollama, which manages models differently.

Adapted from the "Qwen3-7B-Instruct on Windows with LM Studio" lesson of Datawhale's open-source self-llm project^[1], released under the Apache-2.0 License.