GPT OSS 120B is a 117B-parameter Mixture-of-Experts model that activates only 5.1B parameters per token, a design that lets it deliver high reasoning capability while running on a single 80GB GPU such as an NVIDIA H100 or AMD MI300X. OpenAI released it on August 5, 2025 alongside a smaller 21B sibling, and both models ship under the permissive Apache 2.0 license, making them freely usable for experimentation, fine-tuning, and commercial deployment. The model's training drew on reinforcement learning and techniques informed by OpenAI's more advanced internal systems, including o3, which helps explain why the 120B variant achieves reasoning performance close to o4-mini on core benchmarks despite its open-weight footprint.
Beyond raw reasoning, GPT OSS 120B is built for production agentic workflows, offering configurable reasoning effort (low, medium, high), full chain-of-thought access for debugging, native function calling, and strong tool-use behavior that has been independently measured on evaluations like Tau-Bench and HealthBench, in some cases surpassing proprietary peers. It must be used with OpenAI's harmony response format to operate correctly, and a 131k-token context window with a May 2024 knowledge cutoff makes it well suited to long-document analysis and multi-step tool orchestration. Availability has expanded well beyond OpenAI's own distribution, with hosting on platforms like Amazon Bedrock and integration into enterprise tooling such as IBM watsonx Orchestrate, positioning it as a flexible open-weight option for teams building agentic applications on their own infrastructure.
