GPT OSS 120B is the larger of OpenAI's open-source release, distributed under Apache 2.0 and engineered for workloads that need serious reasoning without proprietary restrictions. It is a 117B-parameter Mixture-of-Experts design that activates roughly 5.1B parameters per forward pass, using an MoE routing structure with 128 experts to keep active compute modest. The model targets data-center GPUs and is optimized to fit on a single H100 through native MXFP4 quantization, making it practical for self-hosting and high-volume serving. Its intended use cases lean toward agentic behavior and general-purpose production, supported by native tool use that includes function calling, browsing, and structured output generation, along with full chain-of-thought access for step-by-step reasoning.
Beyond raw capabilities, GPT OSS 120B is shaped around flexibility and control. Developers can tune reasoning effort across low, medium, and high settings to balance latency against answer quality, and the model accepts fine-tuning so teams can adapt it to domain-specific needs. Its long context window supports extended documents and multi-turn agentic sessions, while the open-weight license allows local deployment, modification, and commercial use. The combination of o4-mini-comparable reasoning depth with single-GPU viability makes it especially well suited for organizations that want a transparent, self-hostable backbone for assistants, retrieval-augmented pipelines, or tooling-heavy applications.
