Repolode
Submit
openai

openai/gpt-oss

Open-weight gpt-oss-120b and gpt-oss-20b language models from OpenAI, with reference inference, tool, and client code.

Open-source alternative to:ChatGPT

View on GitHubWebsite
Stars
20,477
Forks
2,179
Open issues
166
Last commit
8 days ago

What is it?

gpt-oss is OpenAI's open-weight model series, released as two models: gpt-oss-120b for high-reasoning use on a single 80GB GPU, and gpt-oss-20b for lower-latency and local use within 16GB of memory. Both models use the harmony response format and should only be used with it. Weights are available on Hugging Face under the Apache 2.0 license.

The repository contains reference implementations in PyTorch, Triton (single GPU, MXFP4 MoE kernels) and Metal (Apple Silicon), along with a terminal chat app, an example Responses API server, and browser and Python tools the models were trained to use. Inference can also run through vLLM, Ollama, LM Studio and Transformers.

The PyTorch and Metal code is intended for education and experimentation rather than production, and the browser and Python tools are sample implementations that need your own safeguards before real deployment.

Key features

  • 01Two open-weight models: 120B for single 80GB GPU, 20B for 16GB memory.
  • 02Adjustable reasoning effort (low, medium, high).
  • 03Function calling, browsing and Python tool use, and Structured Outputs.
  • 04MXFP4-quantized MoE weights and BF16 for other tensors.
  • 05Works with Transformers, vLLM, Ollama and LM Studio.

Use cases

  • Running a capable reasoning model on a single high-memory GPU.
  • Running a smaller model locally on consumer hardware via Ollama or LM Studio.
  • Building agents that call functions or use browsing and Python tools.
  • Fine-tuning an open model for a specific domain.

Quick start

  1. 1

    Install the gpt-oss tools from PyPI

    pip install gpt-oss
  2. 2

    Install Ollama from its download page

  3. 3

    Pull and run the gpt-oss-20b model with Ollama

    ollama pull gpt-oss:20b

Things to know

  • The reference PyTorch implementation is inefficient and needs at least 4 H100 GPUs.
  • The reference Triton implementation requires building Triton from source and a CUDA GPU.
  • The reference implementations are not expected to be run in production.
  • The reference implementations have not been tested on Windows.

Topics

#llm#open-weights#reasoning#pytorch#triton#vllm#mxfp4

FAQ

What is gpt-oss?
Open-weight gpt-oss-120b and gpt-oss-20b language models from OpenAI, with reference inference, tool, and client code. gpt-oss is OpenAI's open-weight model series, released as two models: gpt-oss-120b for high-reasoning use on a single 80GB GPU, and gpt-oss-20b for lower-latency and local use within 16GB of memory. Both models use the harmony response format and should only be used with it. Weights are available on Hugging Face under the Apache 2.0 license.
Is gpt-oss free to use? What license does it have?
gpt-oss is open source under the Apache-2.0 license. Check the license terms before using it in commercial products.
Is gpt-oss still maintained?
Actively maintained. The last commit to gpt-oss was on Oct 12, 2026.
What can I use gpt-oss for?
Running a capable reasoning model on a single high-memory GPU. Running a smaller model locally on consumer hardware via Ollama or LM Studio. Building agents that call functions or use browsing and Python tools. Fine-tuning an open model for a specific domain.
How do I install gpt-oss?
Install the gpt-oss tools from PyPI: pip install gpt-oss Install Ollama from its download page Pull and run the gpt-oss-20b model with Ollama: ollama pull gpt-oss:20b
Where can I get gpt-oss?
The source code and installation instructions are on GitHub: https://github.com/openai/gpt-oss

Summary generated by AI from the project's README. GitHub stats refresh daily.

Related projects