GPT-OSS-120B is OpenAI's large-scale open-weight release, built as a Mixture-of-Experts model with 117 billion total parameters that activates around 5.1 billion parameters per forward pass. This sparsity lets it deliver high-reasoning performance while running efficiently on a single H100 GPU when paired with native MXFP4 quantization. It carries an Apache 2.0 license, so teams can self-host, fine-tune, and adapt it for production use without proprietary restrictions, and the weights are openly available for download and inspection. The model is aimed at demanding production workloads, particularly agentic applications that benefit from step-by-step reasoning and external tool integration. It supports configurable reasoning depth across low, medium, and high effort settings and exposes full chain-of-thought traces, alongside native function calling, browsing, and structured output generation. This combination makes it a strong fit for complex pipelines such as multi-step research assistants, code-generation agents, and retrieval-augmented systems where reliable reasoning and tool orchestration matter more than raw conversational fluency.

Positioned at the top of the GPT-OSS family, GPT-OSS-120B targets data-center-grade deployments rather than consumer hardware, offering reasoning capability comparable to OpenAI's own o4-mini model. Its large context window and expert routing allow it to handle long, structured inputs while keeping latency manageable for high-throughput serving. For organizations that need a transparent, modifiable backbone for advanced reasoning tasks, it offers a practical bridge between frontier proprietary models and fully open-source infrastructure.