How do you run gpt-oss on your own machine?

On August 5, 2025, OpenAI released two open-weight reasoning models, gpt-oss-20b and gpt-oss-120b. Open-weight means the trained model files themselves are free to download, under the permissive Apache 2.0 license, so you can run them on your own hardware instead of sending your text to a company's servers. Running a model locally keeps your data on your machine, works without an internet connection once the model is downloaded, and costs nothing per message.

The two easiest ways to do it are Ollama, a free tool that runs models from the command line and through its own app, and LM Studio, a free desktop application with a chat window. Both worked with OpenAI ahead of the release and supported gpt-oss on launch day.

What you are downloading

Both models use a design called mixture-of-experts: the model is split into many smaller "expert" networks, and only a few are used for each word it processes. That is why the numbers look odd at first. gpt-oss-20b has 21 billion parameters in total but uses about 3.6 billion for each token; gpt-oss-120b has 117 billion in total and uses about 5.1 billion. Fewer active parameters means faster answers than a model of the same total size.

The models also ship already compressed in a format called MXFP4, which stores most of the weights in about 4.25 bits each. Ollama explains that this compression is what lets the smaller model run in as little as 16 GB of memory. Both models handle up to 128,000 tokens of context, are text-only, and were trained mostly on English, with a focus on science, coding and general knowledge.

Option 1: Ollama, step by step

Install or update Ollama from its official website. gpt-oss needs a recent version, since Ollama added new code to handle the MXFP4 format.
Open a terminal (Terminal on a Mac, PowerShell on Windows) and run:

ollama run gpt-oss:20b

The first run downloads the model, which takes a while; later runs start from the local copy.
When the prompt appears, type your question. To test that it really runs locally, switch off your Wi-Fi and ask:

Prompt

Explain in five short bullet points what a mixture-of-experts model is, for someone who has never studied machine learning.

If you have the hardware for the large model, the command is ollama run gpt-oss:120b. Ollama's new desktop app can also download and chat with either model without the terminal.

Option 2: LM Studio, step by step

Download LM Studio from its official website, or update it. Its team asks users to be on version 0.3.21 or later to run gpt-oss.
Use the model search to find gpt-oss-20b and download it.
Load the model and start a new chat.

LM Studio says the smaller model needs around 13 GB of RAM. On Macs with Apple chips, it also runs gpt-oss through MLX, Apple's machine-learning framework, to which the LM Studio team contributed an implementation.

Try it: Install gpt-oss-20b with either tool, turn off your internet connection, and ask it to summarize a document you would not paste into an online chatbot. Note how long the first answer takes.

Using it well

gpt-oss is a reasoning model: it thinks through a problem before answering. You can set its reasoning effort to low, medium or high; OpenAI says this is done with one sentence in the system message, the standing instructions that come before your chat. Low is faster; high is slower but more careful. Ollama gives you access to the model's full chain of thought, its visible working, which can help when you are debugging an answer. OpenAI advises developers not to display that raw reasoning to the people using their apps, because it was trained without direct supervision and may contain mistakes or content not meant for them.

Keep the limits in mind. A local model does not know about recent events unless you give it the information, and like any model it can state wrong facts confidently. Check anything that matters.

Which size fits your hardware?

Use this table as a starting point. The memory figures come from OpenAI, Ollama and LM Studio; real speed depends on your processor, graphics card and what else is running.

Your machine	Model to try	Notes
Less than 16 GB of RAM or GPU memory	Neither, comfortably	Below the memory figures the makers give; try a smaller open model instead
16 GB laptop or desktop, or a 16 GB graphics card	gpt-oss-20b	OpenAI's target for consumer and on-device use; LM Studio cites about 13 GB
32 GB or more	gpt-oss-20b	More headroom for long documents and other apps
A single 80 GB GPU, such as an NVIDIA H100	gpt-oss-120b	The setup OpenAI designed the large model for

For most people, gpt-oss-20b is the right choice. OpenAI says it gives results similar to its o3-mini model on common benchmarks, and it fits on the kind of computer many people already own. Windows users with a suitable graphics card have one more route: Microsoft is offering GPU-optimized versions of gpt-oss-20b in its Foundry Local tool and in the VS Code AI Toolkit.

gpt-oss-120b, which OpenAI says comes close to o4-mini on core reasoning tests, is aimed at servers and serious workstations. If you need its quality without that hardware, the same open weights are offered by cloud hosting providers, though then your data leaves your machine again. Start with the 20b model, see whether its answers meet your needs, and move up only if they don't.