gpt-oss-120b is an open-weight language model released by OpenAI on August 5, 2025, distributed under the Apache 2.0 license with weights hosted on Hugging Face and a companion paper on arXiv. It is a 117-billion-parameter Mixture-of-Experts design that only activates about 5.1 billion parameters per forward pass, and it ships with native MXFP4 quantization so it can run on a single high-memory GPU such as an 80 GB H100. Training drew on reinforcement learning and techniques informed by OpenAI's larger internal systems, including o3, which is the lineage behind its reasoning behavior and chain-of-thought exposure.

In practice, gpt-oss-120b is aimed at production-grade reasoning and agentic workflows, with configurable reasoning depth, full chain-of-thought access, native function calling and browsing, and structured output generation. OpenAI positions it as reaching near-parity with o4-mini on core reasoning benchmarks, and it performed strongly on agentic evaluations such as Tau-Bench and HealthBench, areas where it was reported to surpass some proprietary peers. A 131,072-token context window makes it well suited to long-form analysis, multi-step tool orchestration, and developer pipelines that need open-weight deployment without giving up frontier-style reasoning.