How to Deploy OpenAI’s gpt-oss-120b Model: A Step-by-Step Guide

What is gpt-oss 120B?

gpt-oss 120B is a powerful open-source large language model containing approximately 116 billion parameters. The model matches the performance of OpenAI’s o3-mini and o4-mini on many benchmarks, making it an ideal choice for advanced AI applications. The best part about the model is its flexibility in deployment. You can run it on a single GPU, though four or eight GPUs deliver optimal performance.

Key Features of gpt-oss 120B

The key features of gpt-oss-120b model include:

• 116B Parameter MoE Design: The model uses a Mixture of Experts architecture with 5.1B active parameters per token that ensures efficiency without compromising capability.

• Small Server Deployment: Runs on a single GPU without the need for quantisation, making it highly accessible.

• Extended Context Window: The model also supports a 131K context length, so it can process and understand significantly larger inputs in one go.

If you’re planning to try the latest gpt-oss 120B model, you’re in the right place. Check out our guide below to get started.

Steps to Deploy gpt-oss 120B

Now, let's walk through the step-by-step process of deploying gpt-oss 120B on Hyperstack.

Step 1: Accessing Hyperstack

1. Go to the Hyperstack website and log in to your account.
2. If you're new to Hyperstack, you'll need to create an account and set up your billing information. Check our documentation to get started with Hyperstack.
3. Once logged in, you'll be greeted by the Hyperstack dashboard, which provides an overview of your resources and deployments.

Step 2: Deploying a New Virtual Machine

Initiate Deployment

1. Look for the "Deploy New Virtual Machine" button on the dashboard.
2. Click it to start the deployment process.

Select Hardware Configuration

For gpt-oss 120B GPU requirements, go to the hardware options and choose either from below:

• 4 xNVIDIA RTX A6000
• 1, 2 or 4xNVIDIA H100 PCIe (select any no. of GPUs)

Choose the Operating System

1. Select the "Ubuntu Server 24.04 LTS R570 CUDA 12.8 with Docker". 

Select a keypair

1. Select one of the keypairs in your account. Don't have a keypair yet? See our Getting Started tutorial for creating one.

Network Configuration

1. Ensure you assign a Public IP to your Virtual machine [See the attached screenshot].
2. This allows you to access your VM from the internet, which is crucial for remote management and API access.

Enable SSH Access

1. Make sure to enable an SSH connection.
2. You'll need this to securely connect and manage your VM.

1. Double-check all your settings.
2. Click the "Deploy" button to launch your virtual machine.

Step 3: Accessing Your VM

Once the initialisation is complete, you can access your VM:

Locate SSH Details

1. In the Hyperstack dashboard, find your VM's details.
2. Look for the public IP address, which you will need to connect to your VM with SSH.

Connect via SSH

1. Open a terminal on your local machine.
2. Use the command ssh -i [path_to_ssh_key] [os_username]@[vm_ip_address] (e.g: ssh -i /users/username/downloads/keypair_hyperstack ubuntu@0.0.0.0.0)
3. Replace username and ip_address with the details provided by Hyperstack.

Step 4: Setting up gpt-oss 120B with Open WebUI

1. To set up gpt-oss 120B, SSH into your machine. If you are having trouble connecting with SSH, watch our recent platform tour video (at 4:08) for a demo. Once connected, use the script below to set up gpt-oss 120B with OpenWebUI.
2. Execute the command below to launch open-webui on port 3000.

# 1) Create a docker network
docker network create gpt-oss-net

# 2) Ensure Hugging Face cache directory exists (shared with the container)
sudo mkdir -p /ephemeral/hug && sudo chmod 0777 /ephemeral/hug

# 3) Start vLLM
sudo docker run -d --gpus=all --network gpt-oss-net --ipc=host -p 8000:8000 -v /ephemeral/hug:/hug:rw --name vllm --restart always -e VLLM_ATTENTION_BACKEND=TRITON_ATTN_VLLM_V1 -e HF_HOME=/hug vllm/vllm-openai:gptoss --model openai/gpt-oss-120b --host 0.0.0.0 --port 8000 --async-scheduling --tensor-parallel-size 1 --gpu-memory-utilization=0.95

# 4) Start Open WebUI (points to vLLM's API)
sudo docker run -d --network gpt-oss-net -p 3000:8080 -v open-webui:/app/backend/data --name open-webui --restart always -e OPENAI_API_BASE_URL=http://vllm:8000/v1 ghcr.io/open-webui/open-webui:main

The above script will download and host gpt-oss 120B. See the model card here: https://huggingface.co/openai/gpt-oss 120B for more information.

Please Note: Because the model is large, downloading and hosting can take a few minutes. You should refresh the OpenWebUI page periodically until it shows the model as loaded.

Additional Deployment Notes for gpt-oss 120B

When running gpt-oss 120B on H100 GPUs, make sure to remove the parameter:

• 1 GPU:  --tensor-parallel-size 1
• 2 GPUs:  --tensor-parallel-size 2
• 4 GPUs:  --tensor-parallel-size 4

Following these steps ensures optimal performance and prevents configuration conflicts during deployment.

Interacting with gpt-oss 120B

1. Open your VM's firewall settings.

2. Allow port 3000 for your IP address (or leave it open to all IPs, though this is less secure and not recommended). For instructions, see here.

3. Visit http://[public-ip]:3000 in your browser. For example: http://198.145.126.7:3000

4. Set up an admin account for OpenWebUI and save your username and password for future logins. See the attached screenshot.

And voila, you can start talking to your self-hosted gpt-oss 120B! See an example below.

Running gpt-oss-120b on AI Studio

You can also run gpt-oss-120b inference on AI Studio for as little as $0.10 per 1 million input tokens. Click here to deploy on AI Studio.

 

 

Hibernating Your VM

When you're finished with your current workload, you can hibernate your VM to avoid incurring unnecessary costs:

1. In the Hyperstack dashboard, locate your Virtual machine.
2. Look for a "Hibernate" option.
3. Click to hibernate the VM, which will stop billing for compute resources while preserving your setup.

Why Deploy gpt-oss 120B on Hyperstack?

Hyperstack is a cloud platform designed to accelerate AI and machine learning workloads. Here's why it's an excellent choice for deploying gpt-oss 12B:

• Availability: Hyperstack provides access to the latest and most powerful GPUs such as the NVIDIA H100 on-demand, specifically designed to handle large language models. 
• Ease of Deployment: With pre-configured environments and one-click deployments, setting up complex AI models becomes significantly simpler on our platform. 
• Scalability: You can easily scale your resources up or down based on your computational needs.
• Cost-Effectiveness: You pay only for the resources you use with our cost-effective cloud GPU pricing. 
• Integration Capabilities: Hyperstack provides easy integration with popular AI frameworks and tools.

FAQs

What is gpt-oss 120B?

gpt-oss 120B is an open-source large language model with approximately 116 billion parameters, delivering benchmark results comparable to OpenAI’s o3-mini and o4-mini.

What are the key features of gpt-oss 120B?

 gpt-oss 120B uses a 116B parameter Mixture of Experts design, runs on as few as 2 GPUs without quantisation and supports a 131K context window.

What architecture does gpt-oss 120B use?

The model is built on a Mixture of Experts (MoE) architecture with 5.1B active parameters per token for efficient, high-performance inference.

What is the context length of gpt-oss 120B?

gpt-oss 120B supports a 131K token context window, allowing it to handle very large inputs or extended conversations in one go.

Which GPUs are recommended for running gpt-oss 120B?

For gpt-oss 120B GPU requirements, go to the hardware options and choose either from below:

• 4 xNVIDIA RTX A6000
  
• 1, 2 or 4xNVIDIA H100 PCIe (select any no. of GPUs)

Can gpt-oss 120B run without quantisation?

Yes, it is optimised for small server deployment and can run on as little as 2 GPUs without the need for quantisation.

How does gpt-oss 120B perform compared to commercial models?

gpt-oss 120B performs on par with OpenAI’s o3-mini and o4-mini across multiple benchmarks while being open-source and more deployable.