explaingit

hiyouga/gpt-oss

Analysis updated 2026-07-12 · repo last pushed 2025-08-14

5Audience · developerComplexity · 4/5QuietLicenseSetup · moderate

TLDR

OpenAI's open-weight AI models (120B and 20B) for reasoning and agentic tasks, designed to run on personal hardware. They support tool use, web browsing, and code execution with adjustable thinking depth.

Mindmap

mindmap
  root((repo))
    What it does
      Adjustable reasoning effort
      Exposes thinking process
      Runs on local hardware
    Tech stack
      PyTorch
      Triton
      Apple Metal
    Use cases
      Build support agents
      Local AI experimentation
      Custom AI servers
    Audience
      Developers
      Researchers
      Hobbyists
Click or tap to explore — scroll the page freely

Code map

Detail Auto

An interactive map of this repo's files and how they connect — its source is parsed live in your browser. Click Visualize to build it.

filefunction / class

What do people build with it?

USE CASE 1

Run a capable AI model locally on a high-end GPU or Mac.

USE CASE 2

Build a customer support agent that uses external tools and executes code.

USE CASE 3

Create a custom AI server implementing the OpenAI Responses API.

USE CASE 4

Experiment with chain-of-thought reasoning for debugging and trust.

What is it built with?

PyTorchTritonApple Metal

How does it compare?

hiyouga/gpt-oss1ncendium/aibusteraaronmayeux/ha-hurricane-tracker
Stars555
LanguagePythonPython
Last pushed2025-08-14
MaintenanceQuiet
Setup difficultymoderatemoderateeasy
Complexity4/53/52/5
Audiencedeveloperops devopsgeneral

Figures from each repo's GitHub metadata at analysis time.

How do you get it running?

Difficulty · moderate Time to first run · 30min

Requires a high-end GPU for the 120B model or a capable Mac for the 20B model, and the specific harmony chat format must be applied.

You can freely use, modify, and distribute these models for any purpose, including commercial applications, as long as you include the license notice.

In plain English

gpt-oss is OpenAI's collection of open-weight AI language models designed for reasoning, agentic tasks, and general developer use. The project includes two model sizes: a larger one (120b) aimed at production use cases that can run on a single high-end GPU, and a smaller one (20b) designed for lower latency and local or specialized applications. Both are released under a permissive Apache 2.0 license, meaning developers can freely experiment with and deploy them commercially. At a high level, these models function like other large language models but come with a few built-in advantages. They feature configurable reasoning effort, letting a developer adjust how much the model "thinks" before answering based on whether speed or depth matters more. They also expose their full chain-of-thought reasoning process, which helps with debugging and trust, and they natively support agentic capabilities like calling external functions, browsing the web, and executing Python code. They use a specific chat format called "harmony" that must be applied for the models to work correctly. The target users are developers and researchers who want to run capable AI models on their own hardware rather than relying on a cloud API. For example, a startup building a customer support agent could use the larger model to power tool-using workflows, while a hobbyist wanting to experiment locally on a Mac could use the smaller version. The repository itself provides reference implementations and tools for educational purposes, including code for running the model via PyTorch, Triton, or Apple's Metal framework. It also includes example implementations of the browser and Python code execution tools the models were trained to use. For those who want to get started quickly without diving into the reference code, the models can also be run through popular, user-friendly tools like Ollama or LM Studio, which simplify downloading and chatting with the models on consumer hardware. The project also includes a sample terminal chat application and an example server implementing a basic version of OpenAI's Responses API, making it easier for developers to build their own applications around these models.

Copy-paste prompts

Prompt 1
How do I run the gpt-oss 20b model locally using PyTorch and the harmony chat format?
Prompt 2
Show me how to configure the reasoning effort setting to prioritize speed over depth for the gpt-oss models.
Prompt 3
How do I set up the example terminal chat application to talk to the gpt-oss model?
Prompt 4
Explain how to use the built-in browser and Python code execution tools with these models.

Frequently asked questions

What is gpt-oss?

OpenAI's open-weight AI models (120B and 20B) for reasoning and agentic tasks, designed to run on personal hardware. They support tool use, web browsing, and code execution with adjustable thinking depth.

Is gpt-oss actively maintained?

Quiet — no commits in 6-12 months (last push 2025-08-14).

What license does gpt-oss use?

You can freely use, modify, and distribute these models for any purpose, including commercial applications, as long as you include the license notice.

How hard is gpt-oss to set up?

Setup difficulty is rated moderate, with roughly 30min to a first successful run.

Who is gpt-oss for?

Mainly developer.

Open on GitHub → Explain another repo

This repo across BitVibe Labs

Verify against the repo before relying on details.