The GPT OSS 20B is an open-weight language model engineered to deliver high-level reasoning and agentic capabilities in a compact, efficient package. With 21 billion total parameters and 3.6 billion active parameters, it is specifically optimized for lower latency and local inference, making it suitable for edge devices with as little as 16 GB of memory. The model is designed to support complex agentic workflows, including robust function calling and few-shot tasks, while offering users the ability to adjust reasoning effort levels to balance performance with speed requirements.

Built using a combination of reinforcement learning and techniques derived from advanced frontier systems, this model is trained on the harmony response format to ensure consistent and reliable output. Its design emphasizes transparency through full chain-of-thought access, allowing developers to debug and verify the model's reasoning process. By providing a permissive Apache 2.0 license, the model encourages experimentation and commercial deployment, serving as a versatile tool for developers who need to integrate sophisticated reasoning into specialized or on-device applications without relying on heavy infrastructure.