GPT-OSS 120B marks OpenAI's entry into open-weight reasoning models, built on a Mixture-of-Experts architecture that activates roughly 5.1 billion parameters per forward pass across 36 layers. With 128 specialized experts and Top-4 routing, the model dynamically engages different computational pathways depending on the task at hand. The architecture incorporates Grouped Query Attention and rotary embeddings with RMSNorm pre-layer normalization, positioning it for complex multi-step reasoning while maintaining computational efficiency. Designed primarily for data center–grade GPU deployments, it targets high-volume production workloads and large-scale agentic applications rather than consumer hardware.

Developed with feedback from the open-source community, the model arrives under the Apache 2.0 license, making it freely runnable locally, in the cloud, or as a foundation for fine-tuned variants. It integrates natively with the Responses API and emphasizes strong instruction following, tool use including web search and code execution, and full chain-of-thought visibility. A standout feature is its adjustable reasoning effort—users can toggle between low, medium, and high complexity depending on task demands, balancing quality against latency. This flexibility, combined with structured output support, positions GPT-OSS 120B for demanding applications ranging from autonomous agents and software engineering to research and multilingual assistants.