What are OpenAI's open-weight models?

For years, OpenAI did not release the weights of its major language models after GPT-2, which led to criticism that a company with "open" in its name had become closed. That changed in August 2025, when OpenAI released open-weight models called gpt-oss, which anyone can download and run on their own hardware.

The initial release included two models of different sizes. The larger one was designed to run on a single high-end data centre GPU, while the smaller one could run on a capable desktop or laptop with enough memory. Both are reasoning models, meaning they can think through problems step by step before answering, and they support tool use, such as calling functions and running searches when connected to the right tools.

They were released under the Apache 2.0 licence, one of the most permissive licences, which allows commercial use, modification and redistribution with minimal conditions.

Why did this matter?

1. It signalled a shift in OpenAI's approach, partly in response to strong competition from open models such as DeepSeek, Qwen and Llama.
2. It gave developers access to OpenAI-quality reasoning that can run privately, without sending data to OpenAI's servers.
3. Organisations with strict data rules, such as governments, banks and hospitals, gained another strong option for on-premises AI.

How can you use them?

Download them from Hugging Face or official channels.
Run them locally with tools like Ollama or LM Studio, if your hardware is sufficient.
Use them through cloud providers that host open models.
Fine-tune them for specific tasks.

Typical uses include private document analysis, offline assistants, AI agents that need to run inside a company's own infrastructure, research and custom applications where API costs would be too high.

There are some limits to keep in mind. The open models are not the same as OpenAI's flagship models in ChatGPT, and the most advanced versions still run only on OpenAI's own platform. As with any open model, you are responsible for safety measures, deployment and monitoring when you run them yourself.

OpenAI may release updated open models over time, so check its official announcements and model pages for the latest versions. For developers comparing open models, gpt-oss is worth including in any evaluation.