Run AI Models Without Internet gpt-oss models work locally, making them ideal for private use and on-prem deployments.
Apache 2.0 License: Full Freedom for Developers no restrictions. gpt-oss is truly open.
20B Model for PCs, 120B for Pro Hardware Run gpt-oss-20b on 16GB RAM PCs. The 120B model fits into 80GB GPU setups.
gpt-oss gives developers full visibility into decision-making with chain-of-thought logs.
The models natively support structured outputs, Python execution, and more.
gpt-oss models outperform or match OpenAI’s own o-series on many benchmarks.
Models tested under OpenAI’s Preparedness Framework — comparable to frontier models in safety.
Adopted by AI Sweden, Orange, and Snowflake Early adopters use gpt-oss for secure hosting and fine-tuned, specialized models
gpt-oss marks OpenAI’s return to open weights after 6 years.OpenAI’s First Since GPT-2