GPT-OSS 120B marks a significant departure from OpenAI's traditionally closed approach, representing the company's first open-weight language model released under the Apache 2.0 license since 2019. The architecture is a Mixture-of-Experts design with 117 billion total parameters, but it activates only 5.1 billion parameters per forward pass, making it efficient enough to run on a single 80 GB GPU in production. This MoE structure, combined with native MXFP4 quantization support, positions the model for high-volume reasoning workloads and large-scale deployments without requiring massive infrastructure investments. The model includes configurable reasoning depth and full chain-of-thought access, allowing developers to tune performance versus compute cost depending on the task at hand.

The training approach blends reinforcement learning with techniques distilled from OpenAI's most advanced frontier systems, including o3 and related internal models. This lineage shows up in the model's benchmark performance—it achieves near-parity with OpenAI o4-mini on core reasoning benchmarks while running efficiently on a single GPU. Beyond raw reasoning, the model demonstrates strong tool use capabilities, performing well on the Tau-Bench agentic evaluation suite and HealthBench, where it has outperformed proprietary models including OpenAI o1 and GPT-4o on tool-related tasks. Its compatibility with the Responses API and native function calling support make it well-suited for agentic workflows where models need to take actions, browse information, and generate structured outputs across extended conversations.