gpt-oss

gpt-oss is a pair of OpenAI reasoning models for developers who want to run a local LLM or serve models on their own hardware. The smaller model targets local and specialized applications; the larger one targets general-purpose tasks that need more reasoning.

• Model choice: gpt-oss-20b targets lower latency, while gpt-oss-120b fits on a single 80 GB GPU with the optimized implementation, including NVIDIA H100 or AMD MI300X hardware.
• Tool use: Both models can work with Python execution and a browser tool for searching and reading web pages; the example terminal client makes these tools optional.
• Application integration: Ollama, vLLM and Transformers are supported inference backends, and an example server provides compatibility with basic Responses API use cases.

The models use the Apache 2.0 license and require OpenAI's Harmony conversation format to work correctly. Reference implementations cover CUDA on Linux and Metal on Apple Silicon, though the Metal implementation isn't production-ready. The educational PyTorch implementation requires substantially more GPU hardware than the optimized implementation.

Web browsing requires external access, and the supplied browser tool is for educational use rather than production. The example Responses API server doesn't implement every API feature or event.