gpt-oss-20b

gpt-oss-20b

Writes, reasons and calls tools from text prompts, in an open-weight 21B model from OpenAI that runs within 16 GB of memory. For developers who want a local or specialized model.

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-20b is the smaller of OpenAI's two open-weight gpt-oss models. It has 21B parameters with 3.6B active, and it is aimed at lower latency and local or specialized use. It handles reasoning, function calling and agent work.

Who it suits: It suits developers who want an open model they can run on their own machine, tune for one job or wire into an agent.

Industry | Who uses it | How they use it | Example
Software | Developers | Run a local model behind an app | A chat feature served from your own computer
AI engineering | Agent builders | Use function calls and code execution in an agent | An agent that browses and runs Python
Research | ML researchers | Fine-tune the model for a narrow task | A tuned version for one domain

People who want the largest model can skip it, since the card points to gpt-oss-120b for heavy production use.

Where to use it:

• Hugging Face Transformers, with a pipeline or the Transformers Serve web server
• vLLM, which starts an OpenAI-compatible server
• PyTorch and Triton reference code, and Ollama, which the card also lists
• Windows, macOS or Linux, where you download the files with Python

Use case | Who | What they get back
Local chat | Developers | Answers from a model that runs within 16 GB of memory
Agent tasks | Agent builders | Function calls, web browsing and Python code execution
Reasoning at a set effort | Researchers | Low, medium or high reasoning effort to match your latency needs

How to set it up: You need these first:

• Python and the huggingface_hub package
• The transformers, kernels and torch packages for the Transformers route
• About 16 GB of memory, which the card says the 20b model fits within

1. Install the Hugging Face hub package and download the model files.
2. Install the packages for the Transformers route.
3. Run the pipeline snippet to get a first reply.

pip install -U huggingface_hub
hf download openai/gpt-oss-20b

pip install -U transformers kernels torch

from transformers import pipeline
import torch

model_id = "openai/gpt-oss-20b"

pipe = pipeline(
    "text-generation",
    model=model_id,
    torch_dtype="auto",
    device_map="auto",
)

messages = [
    {"role": "user", "content": "Explain quantum mechanics clearly and concisely."},
]

outputs = pipe(
    messages,
    max_new_tokens=256,
)
print(outputs[0]["generated_text"][-1])

How to use it: Send chat messages through the pipeline, or start a server and chat with it. The model needs the harmony response format. The Transformers chat template applies it for you. With model.generate you apply it yourself.

transformers serve
transformers chat localhost:8000 --model-name-or-path openai/gpt-oss-20b

Tool | What it does
transformers serve | Starts an OpenAI-compatible web server
transformers chat | Opens a chat with the served model
vllm serve openai/gpt-oss-20b | Downloads the model and starts a vLLM server

Example: You send the pipeline a message that asks how a hash table works in three sentences. The snippet prints the last message of the result. That is the model's reply, up to 256 new tokens long.

For newcomers and experts: An open-weight model is one whose files you can download and run yourself. A first small thing to try is the pipeline snippet above with a short question of your own.

For experts, the model is a mixture-of-experts design with 3.6B active parameters out of 21B. It was post-trained with MXFP4 quantization of the expert weights, and the card says all its evaluations used that same quantization. You can set reasoning effort to low, medium or high. The card also says the full chain of thought is available for debugging.

You send a chat message
The chat template applies the harmony format
The model reasons at your chosen effort
It may call tools or write code
You read the final reply

Good to know:

Title: Active parameters
gpt-oss-20b: 3.6 B
gpt-oss-120b: 5.1 B

• Size: 20.9B parameters in Safetensors format, with 261 quantized and 573 fine-tuned versions made by other people
• Tools: the card lists function calling, web browsing, Python code execution and Structured Outputs
• License: apache-2.0
• Cost to run: none named, no API key or account
• Data it reaches: it downloads model files from Hugging Face, and the card lists web browsing and Python code execution as tools it can use
• Not for: use without the harmony format, which the card says breaks it, or showing its chain of thought to end users

On Hugging Face, 13 November 2026

• Task: text generation
• Library: transformers
• Parameters: 20.9B
• Weights: about 20.9 GB in U8
• Formats: Safetensors
• Downloads in the last 30 days: 5,685,450
• Likes: 5,299
• License: apache-2.0
• Quantized versions: 263
• Fine-tunes: 574
• Tags: transformers, safetensors, gpt_oss, text-generation, vllm, conversational, 8-bit, mxfp4

Links

• Model: https://huggingface.co/openai/gpt-oss-20b
• Files: https://huggingface.co/openai/gpt-oss-20b/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download openai/gpt-oss-20b

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