OpenAI gpt-oss
OpenAI's first open-source release since GPT-2 — 120b/20b, rivals o4-mini; historic
OpenAI's first open-source release since GPT-2: 120b/20b with performance rivaling o4-mini — a historic event.
Open foundation models — the base layers everything builds on.
Key features
- Open-weight reasoning models in 120B and 20B variants
- Permissive Apache 2.0 license; MXFP4 quantized for single-GPU inference
- Configurable reasoning effort; full chain-of-thought access
- Native function calling, web browsing, Python execution, structured outputs
- Runs with Transformers, vLLM, Triton, and Metal implementations
Good for
- Agentic applications with tool use
- Local reasoning LLM deployment
Stack & tags
PythonPyTorchTritonllmopen-weightsreasoningagenticapache-2.0moequantized
Quick start
vllm serve openai/gpt-oss-20bFrom the project README on GitHub.