GPT-OSS is a series of open-weight language models released by OpenAI for reasoning, agentic tasks, and developer use cases.
The project addresses the need for capable open models that can run on accessible hardware while maintaining strong reasoning performance. It solves this through two size variants: a 120-billion-parameter model designed to fit on a single 80GB GPU, and a 20-billion-parameter model for lower-latency or local deployment. Both models use MXFP4 quantization of mixture-of-experts weights to achieve this efficiency. The models are trained with the harmony response format, which structures outputs to include full chain-of-thought reasoning alongside final answers, and they require this format to function correctly.
Developers should choose these models if they need open-weight alternatives with native support for agentic capabilities including function calling, web browsing, Python code execution, and structured outputs. The 120B variant suits production workloads requiring strong general reasoning on capable hardware, while the 20B variant targets scenarios prioritizing latency or resource constraints. Both models are fully fine-tunable and released under a permissive Apache 2.0 license without copyleft restrictions. The project provides reference implementations for PyTorch, Triton, and Metal, alongside integration examples using popular inference frameworks like Transformers and vLLM.
The repository includes comprehensive documentation covering model setup, download procedures, and multiple inference paths. Developers can adjust reasoning effort across low, medium, and high settings to balance latency and output quality for their specific use case. The project maintains active integration guidance with established inference libraries and provides tooling for the harmony format to simplify adoption.
