The gpt-oss repository contains OpenAI's open-weight language models, specifically gpt-oss-120b and gpt-oss-20b, designed for reasoning tasks, agentic applications, and general developer use cases. The repository is written in Python and serves as the primary distribution point for these models along with reference implementations and tooling for their deployment and use.
The repository releases two model variants with different scale and performance characteristics. The gpt-oss-120b model contains 117 billion parameters with 5.1 billion active parameters and is designed for production and general-purpose high-reasoning use cases that fit within a single 80GB GPU such as NVIDIA H100 or AMD MI300X. The gpt-oss-20b model contains 21 billion parameters with 3.6 billion active parameters and targets lower latency scenarios and local or specialized deployments. Both models were trained using OpenAI's harmony response format and require this format for correct operation.
The models are released under a permissive Apache 2.0 license, enabling free use without copyleft restrictions or patent concerns. Key features include configurable reasoning effort levels that can be adjusted based on specific use cases and latency requirements, full chain-of-thought reasoning that provides complete access to the model's reasoning process for debugging and transparency purposes, fine-tuning capabilities for customization to specific applications, and native agentic capabilities including function calling, web browsing, Python code execution, and structured outputs. Both models employ MXFP4 quantization of mixture-of-experts weights, allowing gpt-oss-120b to run on a single 80GB GPU and gpt-oss-20b to operate within 16GB of memory.
The repository provides multiple reference implementations for inference. A PyTorch implementation offers an inefficient but educational version showing the exact model architecture with tensor parallelism support for the larger model. A Triton implementation provides optimization for single-GPU deployment using an optimized Triton mixture-of-experts kernel that supports MXFP4 quantization. A Metal implementation targets Apple Silicon hardware, though it is not production-ready. The repository also includes integrations with popular inference frameworks including Transformers, vLLM, Ollama, and LM Studio.
Beyond inference implementations, the repository includes reference implementations of system tools the models were trained on, specifically a browser tool and a Python code execution tool. Client examples demonstrate practical usage patterns, including a terminal chat application that combines the inference implementations with the available tools and a Responses API compatible server implementation supporting multiple inference backends.
According to GitGenius tracking data, the repository shows median issue and pull request response latency of 626.6 hours with a mean of 736.1 hours across tracked items. The most active contributors tracked include dkundel-openai with four events, followed by Sahil3378 and rjarun8235 with one event each. The repository shares overlapping contributors with simonw/datasette, nomic-ai/gpt4all, and anthropics/claude-code, indicating cross-pollination within the open-source AI development community.
The repository is classified across multiple domains including Language Models, Generative AI, Transformers, Natural Language Processing, AI Tools, Machine Learning, Deep Learning, Text Generation, AI Development, and Foundation Models. Setup requirements include Python 3.12 and platform-specific dependencies, with CUDA required for Linux reference implementations. The models and converted weights are available through the Hugging Face Hub, and the repository provides installation options through PyPI for direct use or local setup for modification and experimentation.
