The smaller model in the gpt-oss family is built for lower-latency, local, and specialized workloads. It has 21 billion total parameters with 3.6 billion active at a time, and the series was trained around OpenAI’s Harmony response format, which the model card says is required for correct operation.

This model is intended for reasoning, tool-oriented work, and agentic applications. Its reasoning effort can be set to low, medium, or high to balance output depth and response time, while parameter fine-tuning supports customization. The Apache 2.0 license also makes it suitable for experimentation, commercial deployment, and other projects that benefit from an adaptable, locally runnable foundation.