gpt-oss-120b

gpt-oss-120b

An open-weight OpenAI language model with 117B parameters, adjustable reasoning effort, visible chain of thought and tool use. Runs on one 80 GB GPU and ships under the Apache 2.0 license.

Made by openai on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: gpt-oss-120b is the larger of two open-weight models in OpenAI's gpt-oss series. It has 117B parameters with 5.1B active at a time, and it is aimed at general-purpose, production and high-reasoning work. A smaller sibling, gpt-oss-20b, is meant for lower latency and local use. Both were trained on OpenAI's harmony response format and need it to answer correctly.

Who it suits: Developers and teams who want a strong open-weight model they can host, tune and use commercially, and who have a large GPU or a server to run it.

Industry | Who uses it | How they use it | Example
Software development | Engineers | Add a self-hosted reasoning model to products | A coding assistant on private servers
Enterprise IT | Platform teams | Serve an OpenAI-compatible endpoint in house | A company chat service that keeps data inside
Research | ML researchers | Fine-tune and study an open-weight model | Adapting it to a narrow field
Agent building | Agent developers | Use its function calling and tool skills | A bot that calls tools to finish a task

If you only have a laptop, you can skip it, because the card sizes it for a single 80 GB GPU. The smaller gpt-oss-20b is the card's pick for local use.

Where to use it:

• Python with the Transformers library, or its built-in server
• vLLM, which can serve an OpenAI-compatible endpoint
• PyTorch and Triton reference code, and Ollama, as the card lists

Use case | Who | What they get back
Reasoning and chat | Developers | Text answers with a reasoning trace you can read
Tool use | Agent builders | Function calls, web browsing, Python runs and structured outputs
Fine-tuning | Researchers | A custom model trained on your own data

How to set it up:

• Python and pip
• About 117 GB of disk space for the weights
• A single GPU with 80 GB of memory, such as an H100 or MI300X, per the card

1. Install the Hugging Face hub tools.
2. Download the model files.
3. Install Transformers, then load the model through a text-generation pipeline, or serve the model with vLLM.

pip install -U huggingface_hub
hf download openai/gpt-oss-120b
pip install -U transformers kernels torch

How to use it: Load the model through a Transformers pipeline and send chat messages. The Transformers chat template applies the harmony format for you. If you call generate directly, apply that format yourself or use OpenAI's harmony package. The pipeline takes the model name, automatic data type and device placement, then a list of chat messages and a token limit, and returns the generated text.

To serve it with vLLM, install the build the card names and start the server.

vllm serve openai/gpt-oss-120b

Tool | What it does
Transformers | Runs the model in Python and applies the harmony format through its chat template
vLLM | Starts an OpenAI-compatible server for the model
PyTorch and Triton | Reference code for running it
Ollama | Another way to run it, which the card lists

Example: You send the pipeline a message asking for a clear, short explanation of quantum mechanics. The model replies as text, and the last item in the result is its answer.

For newcomers and experts: If you are new to this, start with the smaller gpt-oss-20b, which needs far less memory. You can set the reasoning effort to low, medium or high to trade speed against depth.

The weights of the mixture-of-experts layers were post-trained in MXFP4 quantization, which is how the 120b model fits on one 80 GB GPU. The card says all its evaluations used that same quantization. The full chain of thought is visible for debugging, but the card says it is not meant to be shown to end users.

You write chat messages
The harmony format wraps them
The model reasons at your chosen effort
It answers or calls a tool

Good to know:

• Format: the model works correctly only with the harmony response format
• Reasoning: choose low, medium or high effort for each use
• Variants: the Hugging Face record lists 136 quantized and 127 fine-tuned versions made by others

Title: Parameters of the two gpt-oss models, in billions
gpt-oss-120b total: 117 billion
gpt-oss-20b total: 21 billion
gpt-oss-120b active: 5.1 billion
gpt-oss-20b active: 3.6 billion

• License: apache-2.0, with no copyleft limits
• Cost to run: no API key or paid service is named for self-hosting, but the model needs a large GPU
• Data it reaches: downloading fetches model files, and the model reads the text you give it. The card says its browsing and Python tools can reach further when you turn them on
• Not for: showing its raw chain of thought to end users, or running on a small machine

On Hugging Face, 13 November 2026

• Task: text generation
• Library: transformers
• Parameters: 117B
• Weights: about 117 GB in U8
• Formats: Safetensors
• Downloads in the last 30 days: 2,435,242
• Likes: 5,509
• License: apache-2.0
• Quantized versions: 136
• Fine-tunes: 127
• Tags: transformers, safetensors, gpt_oss, text-generation, vllm, conversational, 8-bit, mxfp4

Links

• Model: https://huggingface.co/openai/gpt-oss-120b
• Files: https://huggingface.co/openai/gpt-oss-120b/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download openai/gpt-oss-120b

Then follow the usage in its model card. Nothing in this kit runs by itself.