gpt-oss-safeguard-20b

gpt-oss-safeguard-20b

Classifies text against a safety policy you write and shows its reasoning. For trust and safety teams filtering model inputs and outputs, labeling content online or checking it offline.

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-safeguard-20b is a safety reasoning model from OpenAI, built on gpt-oss-20b. You give it a written policy and some text, and it decides whether the text fits the policy. It has 21B parameters, of which 3.6B are active.

Who it suits: Trust and safety teams and developers who need to label or filter text by rules of their own.

Industry | Who uses it | How they use it | Example
AI products | Developers | Filter what goes into and out of a language model | Check chatbot replies against a house policy
Online platforms | Trust and safety teams | Label posts and comments as they arrive | Tag posts that break a posted rule
Content review | Safety practitioners | Label a backlog of content offline | Sort old reports by policy category

If you only need a general chat model, the card points to the regular gpt-oss models instead.

Where to use it:

• Your own GPU, since the 20B model fits in 16 GB of VRAM
• Python programs that load models with the transformers library
• Pipelines that call it for live filtering or for offline labeling

Use case | Who | What they get back
Input and output filtering | Developers | A policy decision for each message
Online labeling | Trust and safety teams | Labels with the reasoning behind them
Offline labeling | Safety practitioners | A labeled set of content

How to set it up:

• Python with pip
• A GPU with about 16 GB of VRAM
• About 21.5 GB of disk space for the weights

1. Install the Hugging Face hub tool.
2. Download the model files.
3. Load them with transformers and format every prompt in the harmony response format.

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

How to use it: You write your policy in plain language and send it with the content to check, in the harmony format. The model will not work correctly without that format. You can set the reasoning effort to low, medium or high, depending on how much latency you can accept. The card links a prompting guide for writing policies.

Example: You write a short policy that defines what counts as spam, then send a comment with it. The model weighs the comment against your policy and returns a decision, and you can read its reasoning to see why.

For newcomers and experts: If you are new, think of it as a reviewer that reads your rulebook before judging each piece of text. Start with a short policy and a few sample messages, and read its reasoning to see if it matches your intent.

For experts, the model interprets a policy you supply rather than a fixed taxonomy. The reasoning effort is adjustable, and the full reasoning is exposed for debugging. The card says that raw reasoning is meant for developers and safety practitioners.

You write a policy in plain language
You send the policy and the text in harmony format
The model reasons over the policy
You read its decision and its reasoning

Good to know: The 120b sibling exists for heavier use, with 117B parameters. The 20b model is the smaller one.

• License: Apache 2.0, as the Hugging Face record lists it.
• Cost to run: no API key or account is named for a downloaded copy. You need a GPU with about 16 GB of VRAM.
• Data it reaches: downloading reaches Hugging Face. The card does not describe the model reaching files, the web or an account.
• Not for: the card says it is meant for safety uses, and recommends gpt-oss for other applications. Its raw reasoning should not be shown to general users.

On Hugging Face, 13 November 2026

• Task: text generation
• Library: transformers
• Parameters: 21.5B
• Weights: about 21.5 GB in U8
• Formats: Safetensors
• Downloads in the last 30 days: 24,325
• Likes: 293
• License: apache-2.0
• Based on: openai/gpt-oss-20b
• Quantized versions: 14
• Fine-tunes: 7
• Tags: transformers, safetensors, gpt_oss, text-generation, vllm, conversational, 8-bit, mxfp4

Links

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

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