OpenAI Harmony

9 months ago

OpenAI's gpt-oss models use the harmony response format for conversation structures, reasoning output, and function calls.
The harmony format mimics the OpenAI Responses API and supports multiple channels for chain of thought, tool calling, and structured outputs.
The format is essential for gpt-oss models to function correctly and is handled automatically by APIs like HuggingFace, Ollama, or vLLM.
A library is available for working with harmony format models, offering consistent formatting, fast performance, and Python support.
The library can be installed via pip or uv and includes Rust core for performance, with Python bindings.
Example code snippets show how to use the library in both Python and Rust for rendering and parsing conversations.
The project structure includes Rust and Python components, with tests ensuring parity between implementations.
Development setup requires Rust, Python ≥ 3.8, and tools like maturin for building the project.

Hasty Briefsbeta