gpt-oss-120B is OpenAI's larger entry in its first open-weight model pair, released under the Apache 2.0 license so it can be self-hosted, fine-tuned, or commercialized rather than locked behind an API. It is built as a mixture-of-experts transformer with routing across 128 experts, allowing only a subset of weights to activate per token so a 120B-scale model can still serve efficiently. Training leaned on reinforcement learning and techniques informed by OpenAI's internal frontier systems, including o3, which is the lineage behind its reasoning behavior rather than a standard pretrained-only stack. The release positions gpt-oss-120B against OpenAI o4-mini in core reasoning benchmarks, while running on a single 80 GB GPU for accessible deployment.

In practice the model is aimed at data-center workloads, high-volume inference, and agentic pipelines that need reasoning, configurable reasoning effort across low, medium, and high settings, and strong tool use. Evaluation evidence highlights competitive few-shot function calling and chain-of-thought results, including Tau-bench agentic tasks and HealthBench where it edged past some proprietary models. It pairs naturally with fine-tuning for domain-specific assistants and with structured-output workflows, fitting teams that want frontier-style reasoning without proprietary lock-in. For organizations that prefer open licensing and on-premises control but still want near-frontier reasoning quality, gpt-oss-120B sits in a sweet spot between smaller local models and closed flagship APIs.