openai /
gpt-oss-120b
OpenAI's open-weight 120B reasoning model for chat, tool use and agents, Apache 2.0
- 5.4k likes
- 2.9M downloads a month
- 117B parameters
- Advanced
- Text & code
gpt-oss-120b is the larger of two models in OpenAI's gpt-oss series (the other is gpt-oss-20b). The card describes it as meant for production, general-purpose, high-reasoning uses. It has 117B parameters but only 5.1B are active at a time, and it was post-trained with MXFP4 quantization of the MoE weights so it can run on one 80GB GPU such as an NVIDIA H100 or AMD MI300X.
What sets it apart is that you can choose how hard it thinks (low, medium or high), see its full chain-of-thought, and fine-tune it. The limits are practical: it needs the harmony response format and will not work correctly without it (the Transformers chat template applies it for you). An 80GB GPU is not a laptop, and the card says the chain-of-thought is not meant to be shown to end users.
What you need to run it
117B parameters. The weights take about 234 GB at full 16-bit precision (fits a multi-GPU server), or about 71 GB as a 4-bit download, which runs on a 128 GB Mac or a GPU with 128 GB. Long inputs and images need more memory on top.
What it does well
Configurable reasoning effort
You pick low (fast, general dialogue), medium (balanced) or high (deep, detailed analysis) by writing something like "Reasoning: high" in the system prompt.
Full chain-of-thought
You can see how the model reasoned, which helps with debugging and trust. The card says this is not intended to be shown to end users.
Agentic tools
It supports function calling, web browsing, Python code execution and Structured Outputs.
Fits on one 80GB GPU
MXFP4 quantization of the MoE weights lets the 120b model run on a single 80GB GPU. Evals in the card were run with the same quantization.
Fine-tunable and Apache 2.0
It can be fine-tuned on a single H100 node, and the permissive licence allows experimentation, customization and commercial deployment.
What you could build
- Build a chatbot or assistant that lets users trade speed against depth by changing the reasoning level.
- Make an agent that calls your own functions defined with schemas.
- Write a small tool that browses the web or runs Python code through the model's built-in tools.
- Fine-tune the model on a specialised dataset if you have access to an H100 node.
- Serve it behind an OpenAI-compatible web server with vLLM or Transformers Serve and point existing apps at it.
Get started
Install the Python packages, then run the model with the Transformers pipeline snippet from the card. If you would rather try it through a local tool, the card also shows Ollama and LM Studio commands.
pip install -U transformers kernels torchollama pull gpt-oss:120bollama run gpt-oss:120blms get openai/gpt-oss-120bCommands are copied from the model card. Read it on Hugging Face for the full setup.
