OpenAI releases gpt-oss open-weight AI models that can run on PCs locally

The gpt-oss-20b and 120b models can be downloaded and run locally, with no internet required - offering full transparency, agentic functions, and fine-tuning support

OpenAI has released two new open-weight AI models, gpt-oss-20b and gpt-oss-120b. The former model is medium-sized and modest, meaning it can run on a PC with 16GB of memory, whereas the latter is a large-sized model that requires at least 60GB of virtual RAM or unified memory. Both versions of the model are now available for free download through platforms such as Hugging Face, Databricks, Azure, and AWS, under the Apache 2.0 license — permitting broad modifications for commercial use. Notably, this is the first time that OpenAI has released an open-weight model since 2019.

 

These open-weight models can be downloaded and run on computers with the aforementioned specifications. Users do not need to have an active internet connection to use these AI models as access to their provider’s server is not involved. This also lets developers build custom tools using these models.

 

gpt-oss-20b and gpt-oss-120b: Highlights

• Permissive Apache 2.0 license: Use, adapt, and deploy freely without copyleft obligations or patent concerns — well-suited for customisation, experimentation, and commercial use.
• Adjustable reasoning effort: Configure the model’s reasoning depth (low, medium, or high) to match your latency constraints and application needs.
• Transparent chain-of-thought: Full visibility into the model’s step-by-step reasoning, useful for debugging and validation — though not designed for end-user display.
• Supports fine-tuning: Tailor the model’s performance to your domain by fine-tuning its parameters.
• Built-in agent-like functions: Leverage native support for structured output, function calls, Python execution, and web Browse.
• Native MXFP4 quantisation: The models are trained with native MXFP4 precision for the MoE layer, making gpt-oss-120b run on a single H100 GPU and the gpt-oss-20b model run within 16GB of memory.

What are open-weight language models?

Open-weight AI models refer to language models whose trained parameters — known as weights — are made publicly accessible. These weights govern how the model interprets and generates outputs. By releasing them, developers, researchers, and organisations can download and operate the models locally, without needing to rely on external APIs or cloud infrastructure. That said, such models often come with usage licenses that may limit how they can be modified or used commercially.

How are they different from other models?

They occupy a space between fully open-source and entirely closed models. While open-source models typically provide access to both the model weights and source code with few restrictions, open-weight models usually release only the weights and may enforce terms around reuse or monetisation. In comparison, closed models like Google’s Gemini or OpenAI’s GPT-4 keep both their weights and source code private, making them accessible only via paid platforms or APIs.