gpt-oss

advancedspreading

gpt-oss is OpenAI's freely downloadable family of language models, which you can run on your own machine or server without the company's cloud.

In plain terms

You use ChatGPT like a restaurant: you order, the kitchen cooks it, and the food is brought out, but you cannot see into the kitchen. gpt-oss, by contrast, is like taking the chef's trained kitchen robot home: it cooks for you in your own kitchen, but you do not know exactly how and from what it was trained. The model's 'weights', or the billions of numbers formed during training, can be downloaded by anyone.

Two sizes were released: the larger gpt-oss-120b is for a powerful server or high-end workstation, while the smaller gpt-oss-20b runs on a good laptop or desktop. Both work with text, are step-by-step "thinking" (reasoning) models, and can call external tools, so they can also be used as engines for agents.

The main difference compared to standard OpenAI services is that if you run it on your own machine, your data never leaves it, there is no usage-based fee, and you can further train the model for your specific task. In return, the hardware, electricity costs, maintenance, and security are your responsibility.

How it works

You download the model weights from a public model repository (such as Hugging Face) or use a local runner like Ollama or LM Studio.
The runtime loads the model into your computer's memory; the smaller version requires approximately 16 GB of memory, the larger about 80 GB.
You give it a question or task, and the model generates the answer locally, without internet connection, token by token.
You can set how much the model should 'think' before responding: longer reasoning may yield better results but is slower.
If needed, you add tools (search, code execution, own database) or fine-tune it on your own texts, forming the basis of an agent.

An everyday example

A law firm wants to summarize confidential contracts but cannot send them to an external cloud. The system administrator installs gpt-oss on the firm's own server, so colleagues chat with it just like with ChatGPT, while the documents remain in-house.

Who is affected

developerYou receive a free, downloadable, and modifiable model on which you can build your own application or agent without paying for every request. → Code

entrepreneur, executiveSensitive corporate data can be kept in-house, and the cost shifts from usage fees to hardware and operations.

lawyer, citizenFrom a data protection perspective, the situation is different if personal data are not transferred to a foreign provider's servers. → Law

parent, student, teacherAfter downloading, it can be tested and studied on a more powerful machine without internet or subscription to see how a language model works. → Education & Family

A common misconception

Mythgpt-oss is a free, downloadable copy of ChatGPT and is fully open source.

In realityAn independent model family with less knowledge than top models, and only its weights are open: OpenAI did not publish the training data or the full training process.

Limits and risks

It works only with text: it does not understand or generate images or audio.
It lags behind OpenAI's most powerful cloud-based models in knowledge, can make mistakes or fabricate facts just as easily, and may perform worse in Hungarian than in English.
The larger variant requires expensive hardware, while the smaller one may run slowly on weaker machines.
If you run it yourself, built-in protections can be bypassed or disabled; you are responsible for secure use, updates, and preventing abuse.

Similar, but not the same

ChatGPT – ChatGPT is OpenAI's ready-made service running in the cloud, while gpt-oss is a downloadable model that you must run yourself.
open-source model – In the strict sense, open-source AI would require not only the weights but also the training code and detailed descriptions of the training data; in the case of gpt-oss, only the final weights and the code needed for execution are freely available.
Llama – Llama is Meta's open-weight model family with its own, stricter license, while gpt-oss comes from OpenAI and was released under the permissive Apache 2.0 license.

Related concepts

Broader

open-weight model, large language model, foundation model

Narrower

gpt-oss-120b, gpt-oss-20b

On the edge

OpenAI, Llama, local execution, fine-tuning, reasoning model

A brief history

OpenAI released the gpt-oss-120b and gpt-oss-20b models under the Apache 2.0 license in August 2025. This was the company's first open-weight language model since the 2019 GPT-2.

Where things stand

Available both in local runtime environments and offerings from major cloud providers, it is used mainly by developers and organizations sensitive to data protection. It faces strong competition from other open-weight model families, including those developed in China.

Status: 2026 November

Also known as

gpt-oss · GPT-OSS · gpt-oss-120b · gpt-oss-20b · OpenAI open-weight model · OpenAI open model · OpenAI open-weight model · open-weight GPT

Sources

This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.