gpt-oss-safeguard-120b

gpt-oss-safeguard-120b

Sorts text against a safety policy that you write, and shows its reasoning for each decision. It suits trust and safety teams that filter or label content.

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-120b is a safety reasoning model from OpenAI, built on gpt-oss. You give it a written policy and some text, and it judges the text against that policy. It also handles basic safety tasks. It has 117B parameters, with 5.1B active, and fits on a single H100 GPU.

Who it suits: It suits trust and safety teams and developers who need to filter or label content by their own rules. It is meant for safety work, not for general chat.

Industry | Who uses it | How they use it | Example
Trust and safety | Policy and moderation teams | Label posts and messages against a written policy | Offline labeling of a batch of reported posts
Software | AI product developers | Filter what goes into and out of a language model | Checking a chatbot's inputs and replies against house rules
Online platforms | Platform safety engineers | Tag content as it arrives | Live labeling of new content under a policy

People who want a general assistant should use the regular gpt-oss models instead.

Where to use it:

• Python with the transformers library on Windows, macOS or Linux
• A server GPU such as a single H100
• Input and output filters around a language model
• Online and offline content labeling pipelines

Use case | Who | What they get back
Input and output filtering | AI developers | A policy decision on each prompt or reply
Online content labeling | Platform teams | Labels on content as it arrives
Offline labeling | Trust and safety teams | Labels on stored content, with the reasoning shown

How to set it up:

• Python on your computer
• A GPU able to hold the model, such as one H100
• About 117 GB of disk for the weights
• A written safety policy of your own

1. Install the Hugging Face hub package.
2. Download the model files.
3. Prepare your policy and send prompts in the harmony format.

pip install -U huggingface_hub

hf download openai/gpt-oss-safeguard-120b

How to use it: Write your policy, then send it with the text to judge. The model works only with OpenAI's harmony response format and does not behave correctly otherwise. You can set the reasoning effort to low, medium or high. The card points to a prompting guide for writing a policy and to the gpt-oss cookbooks for inference.

Example: You write a short policy for a forum and send one reported post with it. The model labels the post against your policy and shows the reasoning behind the label. You can read that reasoning to debug the policy.

For newcomers and experts: Think of the policy as the rulebook you hand the model each time. A first small step is to start with a short rulebook, a few test posts and the low reasoning setting. Read the reasoning to see why each label was chosen.

It interprets your policy at run time, so one model can serve several products without retraining. The raw chain of thought is meant for developers and safety practitioners, and should not be shown to general users. The 20b sibling has 21B parameters, 3.6B active, and lower latency.

You write a safety policy
Policy and content go in, in harmony format
Model reasons about the content
Label and reasoning come back

Good to know:

• License: Apache 2.0.
• Cost to run: none named for the weights. You need a large GPU of your own.
• Data it reaches: it downloads files from the Hugging Face Hub, then judges the text you send it.
• Not for: other applications. The card recommends the gpt-oss models for those, and says the raw reasoning is not for general users.

Title: Total parameters in billions
gpt-oss-safeguard-120b: 117 billion
gpt-oss-safeguard-20b: 21 billion

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: 16,400
• Likes: 128
• License: apache-2.0
• Based on: openai/gpt-oss-120b
• Quantized versions: 6
• Fine-tunes: 2
• Tags: transformers, safetensors, gpt_oss, text-generation, vllm, conversational, 8-bit, mxfp4

Links

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

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