openai gpt-oss: gpt-oss-120b and gpt-oss-20b are two open-weight language models by OpenAI

If you want to learn about or download the previous version (v1.1.0), please click here. We read every piece of feedback, and take your input very seriously.

Install

Download gpt-oss-120b and gpt-oss-20b on Hugging Face A programming framework for agentic AI We are taking agent harness to the next level — enabling multi-agent collaboration, effortless agent team design, and introducing agents as the unit of work interaction. The ultimate space for work and life — to find, build, and collaborate with agent teammates that grow with you.

Both models were trained using our harmony response format and should only be used with this format; otherwise, they will not work correctly. Welcome to the gpt-oss series, OpenAI’s open-weight models designed for powerful reasoning, agentic tasks, and versatile developer use cases. OpenAI has now released the macOS version of the application, and a Windows version 1xbet login will be available later (Introducing GPT-4o and more tools to ChatGPT free users). 该API Key用于转发API，需要将Host改为api.chatanywhere.tech(国内首选)或者api.chatanywhere.org(国外使用)。 Gpt-oss-120b and gpt-oss-20b are two open-weight language models by OpenAI The reference implementations in this repository are meant as a starting point and inspiration.

• We also recommend using BF16 as the activation precision for the model.
• This implementation is not production-ready but is accurate to the PyTorch implementation.
• In this implementation, we upcast all weights to BF16 and run the model in BF16.
• Both models were trained using our harmony response format and should only be used with this format; otherwise, they will not work correctly.
• We also include an optimized reference implementation that uses an optimized triton MoE kernel that supports MXFP4.

Installation

This reference implementation, however, uses a stateless mode. The model was trained to use a python tool to perform calculations and other actions as part of its chain-of-thought. To improve performance the tool caches requests so that the model can revisit a different part of a page without having to reload the page. The model has also been trained to then use citations from this tool in its answers. To control the context window size this tool uses a scrollable window of text that the model can interact with. You can either use the with_browser_tool() method if your tool implements the full interface or modify the definition using with_tools().

This version can be run on a single 80GB GPU for gpt-oss-120b. To run this implementation, the nightly version of triton and torch will be installed. We also include an optimized reference implementation that uses an optimized triton MoE kernel that supports MXFP4.

To enable the python tool, you’ll have to place the definition into the system message of your harmony formatted prompt. As a result the PythonTool defines its own tool description to override the definition in openai-harmony. During the training the model used a stateful tool which makes running tools between CoT loops easier. To enable the browser tool, you’ll have to place the definition into the system message of your harmony formatted prompt. The torch and triton implementations require original checkpoint under gpt-oss-120b/original/ and gpt-oss-20b/original/ respectively.