GPT OSS 120B is a large-scale open-weight language model released by OpenAI as part of the gpt-oss family, designed to push the frontier of accessible reasoning AI. Built on a Mixture-of-Experts architecture with 128 experts and a routing system that activates only the relevant specialists per query, it delivers efficient inference while maintaining broad capability. The model was trained using a combination of reinforcement learning and techniques informed by OpenAI's frontier internal systems, including o3, giving it strong step-by-step reasoning skills suited to complex problem-solving and agentic workflows.
In practical terms, GPT OSS 120B is positioned for developers who need strong reasoning performance without proprietary constraints. OpenAI reports that it achieves near-parity with its own o4-mini on core reasoning benchmarks while remaining efficient enough to run on a single 80 GB GPU, and it demonstrates strong tool use and few-shot function calling, including on agentic evaluations like Tau-Bench and HealthBench. Released under the Apache 2.0 license, the weights are available on Hugging Face and can be freely run locally or in the cloud, modified, fine-tuned, and commercialized, making the model a flexible foundation for domain-specific tuning, on-premises deployments, and high-volume production workloads that demand configurable reasoning effort.
