GPT OSS 120B is a high-performance, open-weight language model built on a scalable Mixture-of-Experts architecture. With 117 billion total parameters and 5.1 billion active parameters per forward pass, it is engineered to deliver robust reasoning and agentic capabilities while remaining efficient enough to run on a single 80 GB GPU. The model is designed for complex, multi-step tasks, offering developers native support for function calling, structured outputs, and configurable reasoning depth. By providing full access to its chain-of-thought process, it allows teams to debug and refine agentic workflows, making it a versatile tool for moving from experimental demos to durable, production-grade systems.

The model was developed using a combination of reinforcement learning and techniques derived from frontier systems, ensuring strong performance on core reasoning benchmarks and agentic evaluation suites. It is trained on the harmony response format, which is essential for its proper operation and integration. Beyond its reasoning strengths, the model is optimized for practical deployment, supporting flexible customization through fine-tuning and providing a permissive Apache 2.0 license for commercial use. Its design emphasizes developer control, allowing users to adjust reasoning effort based on specific latency and accuracy requirements, positioning it as a forward-looking solution for local inference and specialized agentic applications.