gpt-oss

gpt-oss is OpenAI's first open-weight release since GPT-2: two Apache-2.0 mixture-of-experts reasoning models shipped in August 2025. gpt-oss-120b has 117B total and 5.1B active parameters and fits a single 80 GB GPU such as an H100 or MI300X, while gpt-oss-20b (21B total, 3.6B active) runs within 16 GB of memory on consumer hardware. MoE weights are natively quantized to MXFP4 with everything else in BF16, and the models expose configurable reasoning effort at low, medium, or high, plus function calling, structured outputs, and built-in browsing and Python tools. They require the harmony response format to work correctly, and the repository provides educational PyTorch, Triton, and Metal reference implementations along with the tool code, while support landed quickly in Ollama, LM Studio, vLLM, and llama.cpp. Apache-2.0 licensing means no copyleft or usage restrictions for commercial deployment. With over 20k GitHub stars and a dedicated OpenAI cookbook section, gpt-oss rapidly became a default choice for local reasoning and agent experiments.