GPT OSS 120B is the larger sibling in OpenAI's gpt-oss family, released under an Apache 2.0 license that allows local or cloud hosting, modification, fine-tuning, and commercial use. The model was trained using a mix of reinforcement learning and techniques informed by OpenAI's internal frontier systems, explicitly including o3, and the launch positions it as the company's first open-weight language model release since Whisper and CLIP. Its release is accompanied by a technical paper on arXiv (identifier 2508.10925) and a Hugging Face repository that distributes the weights for community use.

Architecturally, GPT OSS 120B is built as a mixture-of-experts model with 128 experts routed per token, enabling efficient inference by activating only the relevant specialists. It is sized for data-center-grade deployment and is designed to run on a single 80 GB GPU, where OpenAI reports near-parity with o4-mini on core reasoning benchmarks and strong performance on tool use and agentic evaluations such as Tau-bench. These characteristics make the model well suited to production reasoning workloads, agentic pipelines with function calling, and fine-tuning for domain-specific applications where organizations want full control over the weights rather than relying on a hosted API.