GPT OSS 120B is the larger member of OpenAI's open-weight gpt-oss family, released under an Apache 2.0 license alongside the smaller gpt-oss-20b variant. It is built as a Mixture-of-Experts model with 117 billion total parameters and 5.1 billion active parameters per pass, using SwiGLU activations and learned attention sinks to keep inference tractable on a single high-end GPU. The model is text-only on both input and output, and was trained with reinforcement learning and techniques informed by OpenAI's internal frontier systems, including o3. On core reasoning benchmarks it is reported to reach near-parity with OpenAI o4-mini, while also performing strongly on tool-use and agentic evaluations such as Tau-Bench and HealthBench.

Designed for production reasoning workloads, GPT OSS 120B offers configurable reasoning effort levels, full chain-of-thought access for debugging, and native function-calling support, making it well suited to agentic pipelines and developer-facing applications. It is available on Cerebras as a production-tier inference model at roughly 3000 tokens per second, with access through pay-as-you-go public endpoints and reserved-capacity dedicated deployments for higher-throughput and SLA-backed use. The combination of sparse activation, permissive licensing, and parameter-level fine-tunability makes it a practical fit for teams that want open-weight reasoning performance with the throughput of dedicated inference hardware.