The gpt-oss-120b is a large-scale mixture-of-experts model engineered to bring high-performance reasoning and agentic capabilities to local and self-hosted environments. With 117 billion parameters and 5.1 billion active parameters, it is designed to balance deep language understanding with the operational efficiency required for production-grade tasks. The model is specifically optimized for complex reasoning, coding, and function calling, providing developers with a versatile tool that can be integrated into agentic workflows while maintaining the flexibility of an open-weight architecture.

Built using a combination of reinforcement learning and techniques derived from frontier systems, the model benefits from a lineage that emphasizes rigorous reasoning and tool use. It is trained on the harmony response format, which is essential for its proper function and performance. The model offers configurable reasoning effort, allowing users to scale performance based on specific latency needs, and provides full access to its chain-of-thought process for easier debugging. By delivering performance that rivals proprietary systems on benchmarks like Tau-Bench and HealthBench, it serves as a robust foundation for developers looking to build, customize, and deploy AI solutions under the permissive Apache 2.0 license.