Overview

• gpt-oss-120B matches o4-mini performance on core reasoning benchmarks while running on a single 80 GB GPU.
• gpt-oss-20B delivers robust reasoning capabilities on devices with just 16 GB of memory, making it suitable for edge and privacy-sensitive environments.
• OpenAI introduced a worst-case fine-tuning protocol with external expert review to simulate malicious biology and cybersecurity scenarios and prevent model abuse.
• OpenAI published evaluation code, prompts, and assessment guidelines to support community testing and safe deployment of the open models.
• Development tools and guides enable integration with platforms such as Hugging Face, vLLM, Ollama, and llama.cpp for diverse use cases.