OpenAI launches open-weight gpt-oss models

OpenAI has launched gpt-oss-120b and gpt-oss-20b, two open-weight language models that you can download, run and adapt on your own infrastructure. It is the company's first release of this kind since GPT-2.

The models come with an Apache 2.0 license, a permissive license that allows you to use, modify and redistribute them, including in commercial products. The difference from OpenAI's usual models is clear: they do not necessarily depend on a cloud API.

Two models for different teams

gpt-oss-120b is designed for demanding reasoning, programming and tool-use tasks. According to OpenAI, it delivers results close to o4-mini on several benchmarks and can run with 80 GB of memory, enough for a single GPU with that capacity.

gpt-oss-20b is smaller and has lower hardware requirements. It can run with 16 GB of memory, making it suitable for local computers, edge devices or smaller servers. In the tests shared by OpenAI, it matches or exceeds o3-mini on several mathematics, health, programming and tool-use tasks.

This does not mean the two models are identical to their proprietary counterparts. The results come from specific evaluations and depend on the task, configuration and hardware. For now, they also do not include the same integrated tool package as the models available through the API platform.

Reasoning, tools and customization

Both models can follow instructions, call functions and work with tools such as web searches and Python code execution. They also support three levels of reasoning effort: low, medium and high. Developers can choose between faster responses and more time to solve complex tasks.

Their context window reaches 128,000 tokens, meaning they can process lengthy documents or conversations in a single pass. They also support structured outputs, which are useful when you need the response in a specific format, such as JSON for feeding another application.

OpenAI has published the weights on Hugging Face and is distributing them already quantized in MXFP4. Quantization reduces the space and memory required to run a model, with a potentially limited loss of precision. The company has also released tools and examples for using them with PyTorch, Apple Metal, Python and Rust.

In practice, this lets you build an internal assistant that processes sensitive data without sending it to an external provider, adapt the model with your own information or test a product without taking on the cost of cloud infrastructure from the start.

The technical design reduces execution costs

Both models use a mixture-of-experts architecture, known as MoE. Instead of activating all their parameters for each word, they select only the relevant portion at every step.

gpt-oss-120b has 117 billion total parameters, but activates around 5.1 billion per token. gpt-oss-20b has 21 billion and activates 3.6 billion. This difference helps explain how large models can run with less memory and computing capacity than their total size might suggest.

OpenAI has also worked with platforms including Azure, Hugging Face, AWS, Ollama, vLLM, LM Studio and OpenRouter, as well as chip manufacturers such as NVIDIA, AMD, Cerebras and Groq. Microsoft will offer an optimized version of gpt-oss-20b for running models locally on Windows.

Openness also brings risks

Publishing the weights makes it possible to inspect and modify the model, but it also makes it easier for someone to adapt it for harmful uses. OpenAI says it tested versions fine-tuned with specialized biology and cybersecurity data to study that risk.

According to its internal and external evaluations, those versions did not reach high levels of dangerous capability within its preparedness framework. The methodology was reviewed by three independent groups. Even so, a model's behavior can change after third parties modify it, so the launch does not resolve the safety question.

OpenAI will also organize a security testing challenge with a $500,000 fund for researchers and developers. The company will later publish a report and an evaluation dataset based on validated issues.

For you, the main consequence is that you no longer have to choose only between installing a less capable open model or paying for a proprietary API. With gpt-oss, a small company, researcher or developer can run an OpenAI model locally and decide how much to customize it. What needs watching now is its real-world performance outside official tests and how the community responds when it starts adapting the model to specific use cases.