GPT OSS 120B (Cerebras)

GPT OSS 120B is OpenAI's open-weight entry in the gpt-oss family, released alongside a smaller sibling under the Apache 2.0 license. It is built as a 117-billion-parameter Mixture-of-Experts transformer that activates roughly 5.1 billion parameters per forward pass, paired with native MXFP4 quantization so the model can be served on a single high-memory GPU. Training drew on reinforcement learning and techniques informed by OpenAI's frontier systems, including o3, which is reflected in the model's near-parity with o4-mini on core reasoning benchmarks and strong showing on agentic evaluations such as Tau-bench and HealthBench where it edged past older proprietary models like o1 and GPT-4o. A technical paper accompanies the release on arXiv.

In practical use, GPT OSS 120B is positioned for high-reasoning, agentic, and general-purpose production workloads that benefit from transparency and self-hosting. It exposes configurable reasoning depth and full chain-of-thought access, alongside native tool use covering function calling, browsing, and structured output, which makes it well suited to multi-step agents, retrieval pipelines, and code or research assistants that need inspectable reasoning. The wide context window supports long documents and extended tool traces, while the open weights let teams fine-tune, audit, and deploy locally without depending on a single hosted endpoint.