GPT-OSS-120B is a 120-billion parameter open-weight language model built on a Mixture-of-Experts architecture, activating 5.1 billion parameters per forward pass to balance capability with computational efficiency. This design enables the model to run on a single 80 GB GPU through native MXFP4 quantization, making large-scale reasoning tasks accessible without requiring massive infrastructure. Released under the flexible Apache 2.0 license, GPT-OSS-120B prioritizes transparency and customizability for developers and organizations that need to inspect, audit, or fine-tune their AI systems. The model was developed to push the frontier of open-weight reasoning models, delivering strong real-world performance at lower deployment costs compared to fully closed systems.

The training approach for GPT-OSS-120B combined reinforcement learning with techniques distilled from OpenAI's most advanced internal systems, including o3 and other frontier models. This lineage enables near-parity with o4-mini on core reasoning benchmarks while excelling at tool use, few-shot function calling, and chain-of-thought reasoning—outcomes validated on agentic evaluation suites like Tau-Bench and HealthBench, where it even surpassed proprietary predecessors. The model is compatible with the Responses API and designed for agentic workflows, supporting function calling, browsing, and structured output generation out of the box. Its availability across major cloud platforms and enterprise systems like IBM watsonx Orchestrate reflects a broader push toward bringing powerful open-weight models into regulated and high-stakes production environments.