Open models by OpenAI
Advanced open-weight reasoning models to customize for any use case and run anywhere.
gpt-oss
Open reasoning models designed to run locally on desktops, laptops, and in data centersâavailable in 120B and 20B parameters.
gpt-oss-safeguard
Open safety reasoning models that support custom safety policiesâavailable in 120B and 20B parameters.
Permissive license
These models are supported by the Apache 2.0 license. Build freely without worrying about copyleft restrictions or patent riskâwhether you're experimenting, customizing, or deploying commercially.
Designed for agentic tasks
Leverage powerful instruction following and tool use within the chain-of-thought, including web search and Python code execution.
Deeply customizable
Adjust the reasoning effort to low, medium, or high. Plus, customize the models to adapt to your use case with full-parameter fine-tuning.
Full chain-of-thought
Access the full chain-of-thought for easier debugging and higher trust in model outputs.
Interactive demo
We built a simple playground where developers can try both models in the browser.
Model performance
gpt-oss-120b | gpt-oss-20b | OpenAI o3 | OpenAI o4-mini | |
Reasoning & knowledge | ||||
MMLU | 90.0 | 85.3 | 93.4 | 93.0 |
GPQA Diamond | 80.1 | 71.5 | 83.3 | 81.4 |
Humanityâs Last Exam | 19.0 | 17.3 | 24.9 | 17.7 |
Competition math | ||||
AIME 2024 | 96.6 | 96.0 | 95.2 | 98.7 |
AIME 2025 | 97.9 | 98.7 | 98.4 | 99.5 |
Advancing safety standards for open models
Safety is foundational to our open models
Comprehensive safety testing
Our partners
Weâre working with leading deployment and hardware companies to offer these models to the open-source community.
Resources
We want to hear from you
Share feedback and feature requests that may guide future open models. Please note: you wonât receive a direct response. For support and discussion, visit our Hugging Face Community.
