What is gpt-oss, and why is OpenAI releasing open-weight models again?

Yesterday, August 5, 2025, OpenAI released two language models that anyone can download, run and modify: gpt-oss-120b and gpt-oss-20b. They are reasoning models, meaning they work through a problem step by step before answering, and OpenAI says they perform close to some of its own paid models.

This is a notable change of direction. OpenAI shared Whisper and CLIP openly in the past, but its chatbots have only been available through its apps and its paid API. The last language model it released with its weights was GPT-2, more than five years ago.

What OpenAI released

Both come under the Apache 2.0 license, one of the most permissive there is: companies can use, change and sell products built on them without paying OpenAI or asking permission. The weights are on Hugging Face, and OpenAI worked ahead of launch with many platforms, including Ollama, LM Studio, llama.cpp, AWS, Azure and Cloudflare, plus chip makers such as NVIDIA and AMD. Microsoft is also bringing an optimized gpt-oss-20b to Windows.

What "open-weight" means

A language model is, at its heart, a very large set of numbers called weights (or parameters). Training adjusts them until the model predicts text well. Releasing the weights means handing over the finished model: you can run it on your own machine, keep your data private, and fine-tune it, which means training it further on your own examples.

"Open-weight" is narrower than fully open source. OpenAI is not publishing the training data, the mostly English, text-only collection used to teach the models. TechCrunch points out that some labs, such as AI2, do release their data, and that copyright lawsuits make that a sensitive choice for OpenAI.

How the models work

Mixture of experts. Each model is a Transformer, the design behind most modern chatbots, built as a mixture of experts. Think of a hospital with many specialists: for each patient, only a few are called in. Here, gpt-oss-120b has 128 "experts" in each layer, and for every token (a word or piece of a word), a small router picks four. That is why only 5.1 billion of its 117 billion parameters do work at any moment, which keeps it fast and cheaper to run.

Compressed numbers. The weights ship in a compact format called MXFP4, which stores each number with fewer bits. That quantization is what lets the larger model fit in 80 GB and the smaller one in 16 GB.

Adjustable reasoning. Like OpenAI's o-series, both models let you choose how hard they think, with three levels: low, medium and high. Higher effort means more thinking and better answers on hard problems, at the cost of speed. Developers set it with one sentence in the system message, the instructions given to a model before the conversation.

Visible chain of thought. The models show their full reasoning, and OpenAI deliberately did not train that reasoning to look tidy, so researchers can study it for signs of misbehavior. OpenAI warns developers not to show it to end users, since it can contain mistakes or content that the final answer would leave out.

Here is a simple way to see reasoning effort in action. Give the same question at low and then high effort:

A train leaves at 9:40 and arrives at 13:05, with a 25-minute stop on the way.
How long was it actually moving? Explain briefly.

At low effort you should get a quick answer; at high effort, a longer line of reasoning before it. Both should reach 3 hours (3 hours 25 minutes in total, minus the stop).

Strengths and limits

OpenAI says gpt-oss-120b matches or beats o4-mini on competition coding, general problem solving and tool use, and does better on health questions and competition math. TechCrunch reports Codeforces coding ratings of 2622 and 2516 for the two models, ahead of DeepSeek's R1 but behind o3 and o4-mini.

The models can call tools such as web search and Python as part of their reasoning. But they are text-only: they cannot read images or produce audio.

They also hallucinate more, meaning they state false things with confidence. On PersonQA, OpenAI's test of facts about people, the two models were wrong on 49% and 53% of questions, against 16% for o1 and 36% for o4-mini. OpenAI expects this from smaller models, which hold less world knowledge.

On safety, OpenAI tested what an attacker could do by deliberately fine-tuning the larger model on biology and cybersecurity data. Even then, it says, the model did not reach its "high capability" danger level. It is also running a red-teaming challenge with a $500,000 prize fund to find new problems.

Why it matters this week

For individuals and small organizations, a capable reasoning model now costs nothing to download and can run privately, on a laptop or on a company's own servers. For the open-model field, which has been led recently by Chinese labs such as DeepSeek, Alibaba's Qwen and Moonshot AI, OpenAI is now a direct participant. Sam Altman framed the release as an open AI stack built in the United States.

How can you run gpt-oss on your own computer?

Start with gpt-oss-20b. You need a computer with at least 16 GB of memory; a recent laptop with that much may do, though speed depends on your hardware. The easiest route is a desktop app that downloads and runs models for you, such as Ollama or LM Studio, both launch partners. Search for gpt-oss in the app, download the 20b version and start chatting. Everything stays on your machine.

If you cannot run it locally, OpenAI's open model playground lets you try both models in a browser, and many cloud services offer them too. The 120b version needs a single high-end 80 GB GPU, so most people will use it through a provider.

Once it is running, keep its limits in mind: double-check facts, especially about people, and use the reasoning-effort setting to trade speed for care when a task is hard.