The gpt-oss-120b is a Mixture-of-Experts model designed to deliver frontier-level reasoning while remaining practical for resource-constrained environments. With 117 billion total parameters and 5.1 billion active parameters per forward pass, the architecture incorporates SwiGLU activations and learned attention sinks to improve inference efficiency. The model supports configurable reasoning effort, allowing users to dial in the depth of thought based on latency needs. Full chain-of-thought transparency is built in, making reasoning steps inspectable rather than opaque. Released under the permissive Apache 2.0 license, it is fully fine-tunable and optimized to run on a single 80GB GPU, making it viable for on-premises or private cloud deployments where data sovereignty is a priority.

Training blended reinforcement learning with techniques informed by OpenAI's most advanced internal systems, including o3 and other frontier models. The approach enabled the model to achieve near-parity with OpenAI o4-mini on core reasoning benchmarks while excelling at agentic tasks. On the Tau-Bench and HealthBench evaluations, gpt-oss-120b outperformed proprietary systems like OpenAI o1 and GPT-4o in tool use and function calling, demonstrating strong few-shot and chain-of-thought capabilities. Built on the harmony response format and compatible with the Responses API, the model is positioned for production agentic workflows, enterprise reasoning tasks, and developers who need an openly deployable foundation model without vendor lock-in.