README.md
| 1 | --- |
| 2 | license: apache-2.0 |
| 3 | pipeline_tag: text-generation |
| 4 | library_name: transformers |
| 5 | tags: |
| 6 | - vllm |
| 7 | --- |
| 8 | |
| 9 | <p align="center"> |
| 10 | <img alt="gpt-oss-120b" src="https://raw.githubusercontent.com/openai/gpt-oss/main/docs/gpt-oss-120b.svg"> |
| 11 | </p> |
| 12 | |
| 13 | <p align="center"> |
| 14 | <a href="https://gpt-oss.com"><strong>Try gpt-oss</strong></a> · |
| 15 | <a href="https://cookbook.openai.com/topic/gpt-oss"><strong>Guides</strong></a> · |
| 16 | <a href="https://arxiv.org/abs/2508.10925"><strong>Model card</strong></a> · |
| 17 | <a href="https://openai.com/index/introducing-gpt-oss/"><strong>OpenAI blog</strong></a> |
| 18 | </p> |
| 19 | |
| 20 | <br> |
| 21 | |
| 22 | Welcome to the gpt-oss series, [OpenAI’s open-weight models](https://openai.com/open-models) designed for powerful reasoning, agentic tasks, and versatile developer use cases. |
| 23 | |
| 24 | We’re releasing two flavors of these open models: |
| 25 | - `gpt-oss-120b` — for production, general purpose, high reasoning use cases that fit into a single 80GB GPU (like NVIDIA H100 or AMD MI300X) (117B parameters with 5.1B active parameters) |
| 26 | - `gpt-oss-20b` — for lower latency, and local or specialized use cases (21B parameters with 3.6B active parameters) |
| 27 | |
| 28 | Both models were trained on our [harmony response format](https://github.com/openai/harmony) and should only be used with the harmony format as it will not work correctly otherwise. |
| 29 | |
| 30 | |
| 31 | > [!NOTE] |
| 32 | > This model card is dedicated to the larger `gpt-oss-120b` model. Check out [`gpt-oss-20b`](https://huggingface.co/openai/gpt-oss-20b) for the smaller model. |
| 33 | |
| 34 | # Highlights |
| 35 | |
| 36 | * **Permissive Apache 2.0 license:** Build freely without copyleft restrictions or patent risk—ideal for experimentation, customization, and commercial deployment. |
| 37 | * **Configurable reasoning effort:** Easily adjust the reasoning effort (low, medium, high) based on your specific use case and latency needs. |
| 38 | * **Full chain-of-thought:** Gain complete access to the model’s reasoning process, facilitating easier debugging and increased trust in outputs. It’s not intended to be shown to end users. |
| 39 | * **Fine-tunable:** Fully customize models to your specific use case through parameter fine-tuning. |
| 40 | * **Agentic capabilities:** Use the models’ native capabilities for function calling, [web browsing](https://github.com/openai/gpt-oss/tree/main?tab=readme-ov-file#browser), [Python code execution](https://github.com/openai/gpt-oss/tree/main?tab=readme-ov-file#python), and Structured Outputs. |
| 41 | * **MXFP4 quantization:** The models were post-trained with MXFP4 quantization of the MoE weights, making `gpt-oss-120b` run on a single 80GB GPU (like NVIDIA H100 or AMD MI300X) and the `gpt-oss-20b` model run within 16GB of memory. All evals were performed with the same MXFP4 quantization. |
| 42 | |
| 43 | --- |
| 44 | |
| 45 | # Inference examples |
| 46 | |
| 47 | ## Transformers |
| 48 | |
| 49 | You can use `gpt-oss-120b` and `gpt-oss-20b` with Transformers. If you use the Transformers chat template, it will automatically apply the [harmony response format](https://github.com/openai/harmony). If you use `model.generate` directly, you need to apply the harmony format manually using the chat template or use our [openai-harmony](https://github.com/openai/harmony) package. |
| 50 | |
| 51 | To get started, install the necessary dependencies to setup your environment: |
| 52 | |
| 53 | ``` |
| 54 | pip install -U transformers kernels torch |
| 55 | ``` |
| 56 | |
| 57 | Once, setup you can proceed to run the model by running the snippet below: |
| 58 | |
| 59 | ```py |
| 60 | from transformers import pipeline |
| 61 | import torch |
| 62 | |
| 63 | model_id = "openai/gpt-oss-120b" |
| 64 | |
| 65 | pipe = pipeline( |
| 66 | "text-generation", |
| 67 | model=model_id, |
| 68 | torch_dtype="auto", |
| 69 | device_map="auto", |
| 70 | ) |
| 71 | |
| 72 | messages = [ |
| 73 | {"role": "user", "content": "Explain quantum mechanics clearly and concisely."}, |
| 74 | ] |
| 75 | |
| 76 | outputs = pipe( |
| 77 | messages, |
| 78 | max_new_tokens=256, |
| 79 | ) |
| 80 | print(outputs[0]["generated_text"][-1]) |
| 81 | ``` |
| 82 | |
| 83 | Alternatively, you can run the model via [`Transformers Serve`](https://huggingface.co/docs/transformers/main/serving) to spin up a OpenAI-compatible webserver: |
| 84 | |
| 85 | ``` |
| 86 | transformers serve |
| 87 | transformers chat localhost:8000 --model-name-or-path openai/gpt-oss-120b |
| 88 | ``` |
| 89 | |
| 90 | [Learn more about how to use gpt-oss with Transformers.](https://cookbook.openai.com/articles/gpt-oss/run-transformers) |
| 91 | |
| 92 | ## vLLM |
| 93 | |
| 94 | vLLM recommends using [uv](https://docs.astral.sh/uv/) for Python dependency management. You can use vLLM to spin up an OpenAI-compatible webserver. The following command will automatically download the model and start the server. |
| 95 | |
| 96 | ```bash |
| 97 | uv pip install --pre vllm==0.10.1+gptoss \ |
| 98 | --extra-index-url https://wheels.vllm.ai/gpt-oss/ \ |
| 99 | --extra-index-url https://download.pytorch.org/whl/nightly/cu128 \ |
| 100 | --index-strategy unsafe-best-match |
| 101 | |
| 102 | vllm serve openai/gpt-oss-120b |
| 103 | ``` |
| 104 | |
| 105 | [Learn more about how to use gpt-oss with vLLM.](https://cookbook.openai.com/articles/gpt-oss/run-vllm) |
| 106 | |
| 107 | ## PyTorch / Triton |
| 108 | |
| 109 | To learn about how to use this model with PyTorch and Triton, check out our [reference implementations in the gpt-oss repository](https://github.com/openai/gpt-oss?tab=readme-ov-file#reference-pytorch-implementation). |
| 110 | |
| 111 | ## Ollama |
| 112 | |
| 113 | If you are trying to run gpt-oss on consumer hardware, you can use Ollama by running the following commands after [installing Ollama](https://ollama.com/download). |
| 114 | |
| 115 | ```bash |
| 116 | # gpt-oss-120b |
| 117 | ollama pull gpt-oss:120b |
| 118 | ollama run gpt-oss:120b |
| 119 | ``` |
| 120 | |
| 121 | [Learn more about how to use gpt-oss with Ollama.](https://cookbook.openai.com/articles/gpt-oss/run-locally-ollama) |
| 122 | |
| 123 | #### LM Studio |
| 124 | |
| 125 | If you are using [LM Studio](https://lmstudio.ai/) you can use the following commands to download. |
| 126 | |
| 127 | ```bash |
| 128 | # gpt-oss-120b |
| 129 | lms get openai/gpt-oss-120b |
| 130 | ``` |
| 131 | |
| 132 | Check out our [awesome list](https://github.com/openai/gpt-oss/blob/main/awesome-gpt-oss.md) for a broader collection of gpt-oss resources and inference partners. |
| 133 | |
| 134 | --- |
| 135 | |
| 136 | # Download the model |
| 137 | |
| 138 | You can download the model weights from the [Hugging Face Hub](https://huggingface.co/collections/openai/gpt-oss-68911959590a1634ba11c7a4) directly from Hugging Face CLI: |
| 139 | |
| 140 | ```shell |
| 141 | # gpt-oss-120b |
| 142 | huggingface-cli download openai/gpt-oss-120b --include "original/*" --local-dir gpt-oss-120b/ |
| 143 | pip install gpt-oss |
| 144 | python -m gpt_oss.chat model/ |
| 145 | ``` |
| 146 | |
| 147 | # Reasoning levels |
| 148 | |
| 149 | You can adjust the reasoning level that suits your task across three levels: |
| 150 | |
| 151 | * **Low:** Fast responses for general dialogue. |
| 152 | * **Medium:** Balanced speed and detail. |
| 153 | * **High:** Deep and detailed analysis. |
| 154 | |
| 155 | The reasoning level can be set in the system prompts, e.g., "Reasoning: high". |
| 156 | |
| 157 | # Tool use |
| 158 | |
| 159 | The gpt-oss models are excellent for: |
| 160 | * Web browsing (using built-in browsing tools) |
| 161 | * Function calling with defined schemas |
| 162 | * Agentic operations like browser tasks |
| 163 | |
| 164 | # Fine-tuning |
| 165 | |
| 166 | Both gpt-oss models can be fine-tuned for a variety of specialized use cases. |
| 167 | |
| 168 | This larger model `gpt-oss-120b` can be fine-tuned on a single H100 node, whereas the smaller [`gpt-oss-20b`](https://huggingface.co/openai/gpt-oss-20b) can even be fine-tuned on consumer hardware. |
| 169 | |
| 170 | # Citation |
| 171 | |
| 172 | ```bibtex |
| 173 | @misc{openai2025gptoss120bgptoss20bmodel, |
| 174 | title={gpt-oss-120b & gpt-oss-20b Model Card}, |
| 175 | author={OpenAI}, |
| 176 | year={2025}, |
| 177 | eprint={2508.10925}, |
| 178 | archivePrefix={arXiv}, |
| 179 | primaryClass={cs.CL}, |
| 180 | url={https://arxiv.org/abs/2508.10925}, |
| 181 | } |
| 182 | ``` |