Install gpt-oss-120b Locally: Step-by-Step Guide for Windows

Want to use a powerful AI like ChatGPT, but without the cloud, without a subscription and without your data ever leaving your desk? With gpt-oss-120b, OpenAI has released an open-weight language model that you can run entirely on your own computer. In this guide, we’ll show you step by step how to install it on Windows using the free app LM Studio. No programming knowledge required.

Key terms in a nutshell

Don’t worry – you don’t need to understand everything to get started. But you’ll come across these terms in the guide:

Why run gpt-oss-120b on the GEEKOM A9 Mega?

Large language models need one thing above all: lots of fast memory. According to LM Studio, gpt-oss-120b takes up around 65GB. Ordinary PCs and most graphics cards don’t have nearly enough memory for that.

The GEEKOM A9 Mega with AMD Ryzen AI Max+ 395 solves the problem with 128GB of unified memory: the CPU and the GPU share the same large pool. You can assign up to 96GB of it to the graphics (AMD) – enough for gpt-oss-120b, with plenty of headroom left for Windows and longer conversations. You can read more about the platform in our guide to AMD Strix Halo.

On top of that, the model itself is cleverly designed: although gpt-oss-120b has around 117 billion parameters, it only activates about 5.1 billion of them per token (“Mixture of Experts”). That’s why it responds quickly despite its size.

Before you start: the checklist

Step 1: Update your AMD graphics driver

The driver is the link between Windows and the graphics processor. An up-to-date driver is the single most important requirement for running the AI quickly.

1. Open the AMD Software: Adrenalin Edition app (type “AMD Software” into the taskbar search box). If it isn’t installed, download the driver from the AMD support page.
2. Click Check for Updates and install the latest recommended version.
3. Restart Windows.
4. Open Task Manager (Ctrl + Shift + Esc), go to Performance and check that AMD Radeon is listed as a GPU.

Step 2: Give the graphics enough memory

By default, the graphics processor only gets a small share of the 128GB. To let the large model run entirely on the GPU, you increase this share once.

1. Open AMD Software: Adrenalin Edition.
2. Go to Performance → Tuning → System and find Variable Graphics Memory.
3. Select Custom and set it to 96GB.
4. Restart your PC. The setting only takes effect after a restart.

Step 3: Install and set up LM Studio

1. Go to lmstudio.ai and download the Windows (x64) version.
2. Run the downloaded file and follow the installation wizard.
3. Open LM Studio once it’s installed.

Choose where models are stored (important – do this before downloading anything): By default, LM Studio saves models to the C: drive. A 65GB model can quickly fill up your system drive.

1. In LM Studio, open My Models (the folder icon in the left sidebar).
2. Change the Models Directory to an SSD with enough free space, e.g. D:\LMStudio\Models.

Step 4: Download gpt-oss-120b

1. Click the magnifying glass (Discover) in the left sidebar.
2. Type gpt-oss-120b into the search box.
3. Select the official entry openai/gpt-oss-120b. Make sure OpenAI is shown as the publisher.
4. Click Download.

Now it’s time to wait. Here’s roughly how long a 65GB download takes on typical UK broadband speeds:

Step 5: Load the model for the first time

The model is now on your SSD. To chat with it, LM Studio needs to load it into memory.

1. Click the speech-bubble icon (Chat) in the left sidebar.
2. Click Select a model to load at the top centre and choose gpt-oss-120b.
3. Turn on Manually choose model load parameters so you can see the settings.
4. Set the values as shown in the table and click Load Model.

The first load takes one to two minutes, depending on your SSD. You’ll then see the model name at the top and a text box at the bottom.

Why not go straight to the maximum context length?

gpt-oss-120b supports a context of up to around 131,000 tokens. But the larger the context, the more memory LM Studio reserves the moment it loads the model. So start with 32,768 and only increase it once everything runs smoothly:

• 32K: standard for chats, emails and short documents
• 64K: for long PDFs, contracts or meeting minutes
• 128K: only when you need it, and after checking in Task Manager that enough memory is free

Step 6: Check that the GPU is really doing the work

The most common beginner mistake: the AI runs, but only on the CPU – up to ten times slower. Here’s how to check in under a minute:

1. Open Task Manager and go to Performance → GPU.
2. Click the small arrow on one of the graphs and switch it to Compute, if available.
3. Type a longer question into the LM Studio chat and watch the graph while the AI answers.

Step 7: Your first chat

Try entering a question like this:

If the AI gives a sensible answer, the installation is complete. Congratulations – you’re now running your own AI completely offline.

Adjusting the reasoning effort

gpt-oss-120b can “think” more or less thoroughly before it answers. In LM Studio, you’ll find this under Reasoning Effort:

Step 8 (optional): Make the AI available to other programs

This step is only relevant if you want to use your local AI in your own scripts, automations or apps. For normal chatting, you can skip it.

LM Studio provides an OpenAI-compatible API, so many programs that normally use ChatGPT can use your local AI instead (LM Studio docs).

1. Click Developer in the left sidebar (the icon with angle brackets).
2. Check that gpt-oss-120b is loaded.
3. Leave the port at 1234 and switch the server on via Status: Running.

Test it in PowerShell (press the Windows key, type “PowerShell” and open it):

If a list with the model name appears, the API is running. You can send your first request like this:

$body = @{
  model = "openai/gpt-oss-120b"
  messages = @(
    @{
      role = "user"
      content = "Describe the GEEKOM A9 Mega in one sentence."
    }
  )
  max_tokens = 200
} | ConvertTo-Json -Depth 5

Invoke-RestMethod -Uri "http://localhost:1234/v1/chat/completions" `
  -Method Post `
  -ContentType "application/json; charset=utf-8" `
  -Body $body

For Python users

First install Python for Windows and tick “Add python.exe to PATH” during installation. Then run this in PowerShell:

py -m pip install openai

Save this script as test_ai.py and run it with py test_ai.py:

from openai import OpenAI

client = OpenAI(
    base_url="http://localhost:1234/v1",
    api_key="lm-studio"
)

response = client.chat.completions.create(
    model="openai/gpt-oss-120b",
    messages=[
        {
            "role": "user",
            "content": "Introduce yourself in three sentences."
        }
    ],
    max_tokens=300,
)

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

Step 9 (optional): Access from other devices on your home network

By default, your AI is only reachable on the A9 Mega itself – the safest setting. If you’d like to use it from a laptop or another PC on the same network:

1. On the Developer page, enable Serve on Local Network.
2. Turn on authentication and set an API key.
3. Allow LM Studio through Windows Firewall for private networks only, not public ones.
4. On the other device, use the address http://IP-OF-YOUR-A9-MEGA:1234/v1, e.g. http://192.168.1.50:1234/v1.

Troubleshooting: the most common problems

FAQs: Running gpt-oss-120b locally

Conclusion

With LM Studio and the GEEKOM A9 Mega, you can run one of the most capable open-weight AI models entirely on your own machine in about half an hour, plus download time. The key points for a smooth start: an up-to-date AMD driver, 96GB assigned to the graphics, GPU Offload set to Maximum, and a 32K context to begin with. That way your AI stays fast and stable – and your data stays with you.