GPT OSS

We built gpt-oss as OpenAI's open-weight models for powerful reasoning and agentic tasks. Two flavors: gpt-oss-120b (117B params, 5.1B active) for production on single 80GB GPU, and gpt-oss-20b (21B params, 3.6B active) for lower latency. Both use our Harmony response format for structured I/O.

Our system provides multiple inference backends (Metal for Apple Silicon, Triton/PyTorch/vLLM for CUDA), a Responses API server, tools for browsing and code execution, and comprehensive evaluation suite. See GitHub for source.

🎯 Key Features

Sparse MoE architecture - 128 experts, 4 active per token
MXFP4 quantization - Efficient weight storage with block scaling
Multiple backends - Metal (Apple), Triton/PyTorch (CUDA), vLLM, Transformers
Built-in tools - Web search (browser), Python execution, patch application
Evaluation suite - GPQA, AIME, HealthBench benchmarks

All components integrate through Harmony encoding for consistent message formatting. See README for setup and Hugging Face for model downloads.