gpt-oss-120b is the larger member of OpenAI's open-weight gpt-oss family, designed for production-scale reasoning and agentic workloads. It uses a Mixture-of-Experts architecture with 117B total parameters but only 5.1B active per forward pass, a sparsity pattern that lets it deliver heavyweight reasoning while keeping inference efficient. Native MXFP4 quantization lets the model run on a single H100 GPU, and additional architectural touches such as SwiGLU activations and learned attention sinks support stable long-context behavior for developer and enterprise deployments.

Practically, the model is tuned for developers who want controllable, transparent reasoning rather than a single fixed behavior profile. Users can dial reasoning depth up or down to match latency budgets, inspect the full chain of thought rather than treating it as a black box, and lean on native tool use for function calling, browsing, and structured output generation. Its text-only design and the Apache 2.0 license make it a strong fit for on-premises or private-cloud deployments where data control matters, while its position as the larger sibling to gpt-oss-20b gives teams a clear upgrade path when they need deeper reasoning without leaving the same model family.