OpenAI has just released two new AI models that you can download and run on your own hardware. These models, called gpt-oss-120b and gpt-oss-20b, are the company’s first open-weight models since GPT-2 in 2019.
Unlike the cloud-based models like ChatGPT, which run on powerful servers, these new models are designed for local use, giving users more control, flexibility, and privacy.
The two models differ in size and hardware requirements. The smaller model, gpt-oss-20b, has 21 billion parameters but uses a technique called Mixture-of-Experts (MoE) to only activate 3.6 billion parameters per token.
This means it can run on a computer with at least 16GB of RAM. The larger model, gpt-oss-120b, has 117 billion parameters, using 5.1 billion per token with MoE, and requires around 80GB of RAM. While this makes it difficult to run on most home PCs, it can run on high-end GPUs like the Nvidia H100. Both models support a context window of up to 128,000 tokens, which allows them to process large chunks of text at once.
A key feature of these models is their support for Chain of Thought (CoT) reasoning, which controls how deeply the model “thinks” before answering. You can adjust the CoT level easily in the system prompt, choosing between low, medium, and high settings. Lower settings are faster but less accurate, while higher settings produce better answers but require more computing power.
In terms of performance, gpt-oss-120b is close to OpenAI’s own o3 and o4-mini cloud models in most tasks. The smaller model, gpt-oss-20b, performs just slightly behind. Both models do especially well in areas like math and coding. However, they aren’t quite as strong in knowledge-based benchmarks like Humanity’s Last Exam, where gpt-oss-120b scored 19% compared to o3’s 24.9%. For comparison, Google’s top model scored 34.8% on the same test.
It’s important to note that these models are text-only and do not support images, audio, or video. OpenAI doesn’t expect them to replace its more powerful GPT models. Instead, these models are meant for situations where users need fast, private, and customizable AI tools running locally — for example, companies that don’t want to send sensitive data to the cloud.
OpenAI knows that many developers already use open-source models for these reasons, but those models usually come from other organizations. Now, with the release of gpt-oss, users can stay within the OpenAI ecosystem while still enjoying the benefits of open, locally controlled AI. The models are released under the Apache 2.0 license, which allows for modification, redistribution, and commercial use.
To ensure safety, OpenAI built strong controls into these models. The company even tried to deliberately tune the models to behave badly, but found they still didn’t perform “evil” tasks well. OpenAI says this is thanks to systems like deliberate alignment and instruction hierarchy, which help prevent misuse.
If you want to try these models, you can download gpt-oss-120b and gpt-oss-20b from HuggingFace, or explore the code and tools on GitHub. OpenAI is also hosting demo versions online.