gpt-oss

by OpenAI · Website

OpenAI's first open-weight release since GPT-2 — a pair of Apache 2.0 Mixture-of-Experts reasoning models shipped natively in MXFP4 quantization. gpt-oss 20B (3.6B active) targets 16 GB consumer hardware with o3-mini-class performance, while gpt-oss 120B (5.1B active) approaches o4-mini and fits a single 80 GB GPU or a high-memory Mac Studio. Both support adjustable reasoning effort and agentic tool use, and rank among the most-pulled models on Ollama.