gpt-oss-20b

gpt-oss-20b is a computer program that writes texts and answers questions. Such programs are called language models: they have learned from vast amounts of text how language works, and assemble answers word by word. The company OpenAI, known for ChatGPT, released gpt-oss-20b in August 2025. The difference to ChatGPT lies not in how it’s used, but in its availability: anyone may download the file containing the fully trained model for free and run it on their own computer. The “oss” part of the name stands for “open source software”, meaning software that is freely distributed. The “20b” refers to roughly 20 billion adjustable numerical values inside the model.

OpenAI’s return to open weights

For years, OpenAI had only rented out its best models over the internet. You sent a question to the company’s servers and got an answer back. The last freely downloadable model before that was GPT-2 from 2019. gpt-oss-20b ended this pause and was therefore something of a small sensation in the industry.

The background is competition. Companies like Meta, Mistral, and especially Chinese providers such as DeepSeek and Alibaba had released strong models for download. Developers worldwide built their tools on top of these models. Whoever doesn’t take part here loses influence over which technology the next generation of programs is built on.

For users, there’s a second important point. If the model runs on your own device, the texts you enter never leave the machine. That’s a real argument for doctors' offices, lawyers, or government agencies. There are also no ongoing costs per query, and the model works even without an internet connection.

20 billion parameters, but only 3.6 active

The 20 billion numerical values in the model are called parameters. They are the result of training and determine which answer comes out. Normally, the rule is: more parameters, more computation per word. gpt-oss-20b gets around this with a trick called Mixture of Experts. The model consists of many small specialized building blocks, and for each word only a few of them are switched on.

Only about 3.6 billion parameters are active per step. The model therefore computes as briskly as a significantly smaller one, but can still draw on the knowledge of all the building blocks. In addition, OpenAI stores the parameters in a very compact numerical format. This means the model fits into about 16 gigabytes of memory, the kind found in a modern gaming PC or a decent MacBook.

One special feature is the adjustable amount of thinking effort. You can tell the model whether it should think briefly or extensively before answering. At high effort, it first writes a kind of scratch pad with intermediate steps, which makes it more accurate on math or programming tasks. This costs time, because more text has to be generated.

From download to company server

The easiest way to encounter gpt-oss-20b is through small programs like Ollama or LM Studio. You install them, select the model from the list, wait for the roughly twelve-gigabyte download, and start typing. The interface looks similar to a chat window in a browser. It’s just that the computation happens on your desk instead of in California.

In companies, the model is more often hidden invisibly inside other products. It answers questions about internal manuals, sorts incoming emails, or suggests code to programmers. Its bigger sibling gpt-oss-120b needs a data-center graphics card for that, whereas the 20b version is often already enough.

A common misconception: open weights do not mean open training data. OpenAI does not reveal which texts the model learned from, nor does it publish the training code. You only get the finished result, but you may freely use it, redistribute it, and further train it for your own purposes. Experts therefore more precisely speak of “open weights” rather than true open-source AI.