gpt-oss: What OpenAI's open model means for companies

gpt-oss is a family of open language models from OpenAI under the Apache 2.0 license; the model card was released in August 2025. The larger gpt-oss-120b runs on a single graphics card with 80 GB of memory, while the smaller gpt-oss-20b gets by with 16 GB. This allows companies to use OpenAI technology on their own hardware or in German data centers without sending data to OpenAI.

For a long time, OpenAI was synonymous with closed models in the US cloud. With gpt-oss, there are open weights from the ChatGPT manufacturer again for the first time. What does this mean for a medium-sized company?

The two models

Both models use a mixture-of-experts architecture: only a small part of the model is active per text module. This makes them faster and more economical than the overall size suggests (according to the model page on Hugging Face).

What gpt-oss can do

• Adjustable thinking effort: low, medium, or high – more thinking for difficult tasks, less for quick answers.
• Tool calls: Calling functions, web search, running Python – important for AI agents.
• Structured outputs: reliable formats such as JSON for automations.
• Customizable: Fine-tuning to your own data is possible.

Three ways to use gpt-oss

1. Via API from Germany: The IONOS AI Model Hub offers gpt-oss-120b for €0.15 input and €0.65 output per million tokens (net, as of October 2026).
2. Locally on a workstation or small server: gpt-oss-20b fits hardware with 16 GB of memory – for tests or simple internal applications, for example.
3. On your own GPU server: gpt-oss-120b for demanding tasks with complete data sovereignty. Hardware for your own AI explains what is needed for this.

What gpt-oss is suitable for in the company

• AI agents and workflows with tool calls, e.g. in n8n
• Document questions in an AI search when medium models are not enough
• Reading data from inquiries, invoices, forms
• internal ChatGPT via Open WebUI

Boundaries

• gpt-oss is not the model behind ChatGPT; the latest closed OpenAI models can be stronger in some tasks.
• The model expects a certain response format ("harmony"); software for operation must support this – common tools do.
• You should test how good the German is for your tasks – see Choosing a language model for German.

License: uncomplicated

Apache 2.0 allows commercial use, adaptation and distribution, as long as license and copyright notices are preserved. This is easier than the Llama license, for example, as the article Open Weight or Open Source? shows.

We use gpt-oss alongside Mistral, Qwen and Llama, depending on what solves a task best – more on the page Open Source Language Models.

Counselling and accompanied implementation is eligible for funding: Through INQA coaching, the federal government covers 80% of the consulting costs of up to €1,200 per day for up to 12 days, i.e. up to €11,520 subsidy. In Mecklenburg-Western Pomerania and the other new federal states, BAFA consulting funding is also eligible at 80% until 31.12.2026 (50% in the old federal states). Funding is provided for consulting, not software. The funding check shows which program fits.

Frequently Asked Questions

What is gpt-oss?

A family of open language models from OpenAI under Apache 2.0, consisting of gpt-oss-120b and gpt-oss-20b. The models can be downloaded and operated by yourself.

What hardware does gpt-oss need?

According to OpenAI, gpt-oss-120b runs on a single graphics card with 80 GB of memory, gpt-oss-20b with 16 GB of memory.

Can I use gpt-oss commercially?

Yes. The Apache 2.0 license allows commercial use, customization, and distribution if the license notices are retained.

Does gpt-oss data go to OpenAI?

No, if you operate the model yourself or use it through a provider in Germany such as the IONOS AI Model Hub. The weights then run completely outside of OpenAI.

Would you like to test gpt-oss with your tasks? Get in touch via the contact page.