gpt-oss-120b: OpenAI’s open reasoning model now on Swiss infrastructure

The Model at a glance:

• OpenAI’s Open-Weight model, Apache 2.0 licence.
• Mixture-of-Experts: 128 experts per layer, 4 active per token.
• Context window of up to 128k tokens.
• Adjustable reasoning effort (low, medium, high).
• Managed operations on Swiss infrastructure.

OpenAI has released gpt-oss-120b, an open-weight model, under the Apache 2.0 licence. We are now offering it as part of our AI on Demand portfolio on Swiss infrastructure, fully managed. This provides businesses with a powerful language model for analysis, coding and agentic workflows, without having to operate the infrastructure themselves.

With the gpt-oss models, OpenAI introduced its first open-weight language models since GPT-2. They were developed using reinforcement learning approaches based on insights from OpenAI’s frontier systems, including o3. For businesses, this means: strong reasoning capabilities, openly licensed and therefore deployable on their own or sovereign infrastructure.

gpt-oss-120b utilises a Mixture-of-Experts architecture: 36 layers, each with 128 experts, of which only 4 are activated per token. The model thus activates approximately 5.1 billion parameters per token, combining high model capacity with efficient inference. The native context length is up to 128k tokens. In addition, the reasoning effort per task can be adjusted across three levels (low, medium, high), which allows for a targeted balance between latency and performance.

The model is suitable for complex tasks that involve a combination of analysis, reasoning, coding and multiple processing steps. Typical areas of application include knowledge assistants, RAG solutions, code analysis and code generation, structured data processing, tool calls and agentic workflows. It also supports structured outputs and tool use, such as web searches or Python code execution, when connected to the appropriate tools and execution environment.

We provide gpt-oss-120b as part of our AI on Demand portfolio: the model runs on Swiss infrastructure, with deployment, monitoring, maintenance, support and personalised advice provided by stepping stone. Model inference takes place in Switzerland, whilst we take care of day-to-day operations. When connecting external tools or services, those integrations determine where any additional processing occurs.

→ Find out more about gpt-oss-120b