What Do You Need to Run gpt-oss Locally? VRAM by Version
gpt-oss currently has 3 versions you can deploy locally. The smallest, gpt-oss 20B, needs 13GB of VRAM at 4-bit; the largest, gpt-oss 120B, is an MoE model and needs 4GB of VRAM plus about 62GB of RAM with MoE offloading. The cheapest device that gets one running is the Tesla V100 16GB, at about $370.
| Version | Released | Context | License | ||
|---|---|---|---|---|---|
| gpt-oss Safeguard 20BMoE | Oct 29, 2025 | 20.9B | 3.6B | 128K | Apache-2.0 |
| gpt-oss 120BMoE | Aug 5, 2025 | 116.8B | 5.1B | 128K | Apache-2.0 |
| gpt-oss 20BMoE | Aug 5, 2025 | 20.9B | 3.6B | 128K | Apache-2.0 |
About gpt-oss
gpt-oss is OpenAI's family of open-weight language models. OpenAI was founded in San Francisco in 2015 and created the GPT series and ChatGPT. Released on August 5, 2025, gpt-oss was the first time OpenAI had published language model weights since GPT-2 in 2019, and it is positioned as an open model you can deploy yourself, aimed at reasoning and agentic tasks. According to the ATOM Report's figures through May 2026, gpt-oss 20B is the most downloaded model on Hugging Face in the 10B to 50B parameter range.
The first two models are gpt-oss 120B (117B parameters, 5.1B active per token) and gpt-oss 20B (21B parameters, 3.6B active per token), both MoE. Their attention layers alternate between looking only at the most recent 128 tokens and looking at the full text, and the context is 128K. The expert weights were already in 4-bit MXFP4 format at the post-training stage. Both are reasoning models with three reasoning effort levels, low, medium and high, and they need to be used with OpenAI's harmony chat format. In October 2025, OpenAI worked with ROOST and other organizations to release gpt-oss Safeguard (120B and 20B), content moderation models fine-tuned from gpt-oss. The whole family is licensed under Apache 2.0, which allows commercial use.
- VendorOpenAI
- Versions3
- Parameters20.9B – 116.8B
- Latest versiongpt-oss Safeguard 20B
- License
- Apache-2.0
- Official site
- openai.com/open-models
- Hugging Face org
- openai
How much VRAM do gpt-oss models need at 2-bit?
2-bit, also called Q2, covers common GGUF files such as Q2_K and UD-Q2_K_XL. They are the smallest and lose the most quality, which makes them a way to get a model running first when VRAM is tight. Each row in the table below is a gpt-oss version and each column is a context length; each cell is the total VRAM needed to deploy it at Q2 (weights + KV cache + 1.1GB runtime overhead).
| Version | WeightsGB | 32K contextGB | 128K contextGB | 256K contextGB | 1M contextGB |
|---|---|---|---|---|---|
| gpt-oss Safeguard 20B20.9B (3.6B active) | 8.3est. | 10.2 | 12.6 | — | — |
| gpt-oss 120B116.8B (5.1B active) | 46.6est. | 48.9 | 52.5 | — | — |
| gpt-oss 20B20.9B (3.6B active) | 8.3est. | 10.2 | 12.6 | — | — |
How much VRAM do gpt-oss models need at 4-bit?
4-bit, also called Q4, corresponds to the widely used Q4_K_M files, which balance size and quality and are the default choice for local deployment. Each row in the table below is a gpt-oss version and each column is a context length; each cell is the total VRAM needed to deploy it at Q4.
| Version | WeightsGB | 32K contextGB | 128K contextGB | 256K contextGB | 1M contextGB |
|---|---|---|---|---|---|
| gpt-oss Safeguard 20B20.9B (3.6B active) | 11.6GGUF | 13.5 | 15.9 | — | — |
| gpt-oss 120B116.8B (5.1B active) | 62.8GGUF | 65.1 | 68.7 | — | — |
| gpt-oss 20B20.9B (3.6B active) | 11.6GGUF | 13.5 | 15.9 | — | — |
How much VRAM do gpt-oss models need at 8-bit?
8-bit, also called Q8, corresponds to Q8_0 files, which come closest to the original in quality and are also the largest. Each row in the table below is a gpt-oss version and each column is a context length; each cell is the total VRAM needed to deploy it at Q8.
| Version | WeightsGB | 32K contextGB | 128K contextGB | 256K contextGB | 1M contextGB |
|---|---|---|---|---|---|
| gpt-oss Safeguard 20B20.9B (3.6B active) | 12.1GGUF | 14 | 16.4 | — | — |
| gpt-oss 120B116.8B (5.1B active) | 63.4GGUF | 65.7 | 69.3 | — | — |
| gpt-oss 20B20.9B (3.6B active) | 12.1GGUF | 14 | 16.4 | — | — |
FAQ
Is gpt-oss open source and free for commercial use?
Yes. gpt-oss 20B, gpt-oss 120B and gpt-oss Safeguard 20B are all licensed under Apache 2.0, so you can use them commercially, fine-tune them and redistribute them, as long as you include the license, keep the copyright notices and mark any files you changed. Each repository also includes a very short usage policy that only asks you to comply with applicable law. There is no monthly-active-user threshold and no separate list of prohibited uses. Safeguard is a content moderation model built on gpt-oss, and it carries the same license as gpt-oss.
Can you run gpt-oss with Ollama?
Yes. The gpt-oss tags are gpt-oss:20b and gpt-oss:120b, published through a partnership between OpenAI and Ollama. Safeguard lives in a separate library under the tag gpt-oss-safeguard:20b, and there is also a gpt-oss-safeguard:120b that this site doesn't list. The -cloud tags, gpt-oss:20b-cloud and gpt-oss:120b-cloud, run on Ollama's servers, not on your own machine.
