--- title: "gpt-oss-20b API hosted in the EU" description: "Run OpenAI's compact open-weight gpt-oss-20b reasoning model on inference.eu: the OpenAI API format, processed in EU data centres, billed in euros." canonical: "https://inference.eu/models/gpt-oss-20b/" language: "en" --- Reasoning · OpenAI # Run gpt-oss-20b in the EU through an OpenAI-compatible API gpt-oss-20b is the smaller of OpenAI's two open-weight reasoning models: 21 billion parameters, 3.6 billion active per token, released under Apache 2.0. On inference.eu you call it with the OpenAI SDK you already use, while your prompts are processed in EU data centres and never used for training. [Contact sales](https://inference.eu/contact/)[Example request](#example) 1. 0121B · 3.6B 2. 02EU data centres 3. 03Apache 2.0 ## What gpt-oss-20b is gpt-oss-20b was released by OpenAI in August 2025 together with gpt-oss-120b. It is a mixture-of-experts model with 21 billion parameters, of which 3.6 billion are active for each token, and a 128K-token context window. Its expert weights were post-trained in MXFP4 precision, so the model fits within 16 GB of memory. Like its larger sibling, it reasons before answering, and you set the reasoning effort per request to low, medium or high. It was trained on OpenAI's harmony response format and supports function calling and structured outputs natively. It reads and writes text only. OpenAI positions it for lower-latency and specialised use cases. ## What it does well, and where it falls short gpt-oss-20b is a quick, inexpensive reasoning model for everyday tasks: answering support questions, extracting fields into JSON, routing requests to tools and handling short analytical questions. With only 3.6 billion active parameters it responds fast, and the adjustable reasoning effort lets you spend thinking only where a request needs it. It is clearly weaker than gpt-oss-120b on hard maths, long code changes and broad world knowledge, and its knowledge is older than 2026 models. It has no image or audio input. OpenAI does not publish a list of supported languages, so check quality in German, French or other European languages against your own content first. ## Calling gpt-oss-20b through the API If your code uses the OpenAI SDK, change the base URL to the inference.eu API, use your inference.eu key and set the model to `gpt-oss-20b`. The harmony format is handled on our side, so you send and receive ordinary chat messages. Pass `reasoning_effort` as `low` for fast replies or `high` for careful answers. Function calling, JSON schema outputs and streaming work as in OpenAI's own API. The chain of thought is intended for debugging, not for end users, so show only the final answer in your interface. The Anthropic Messages format is also available if your tools were built for Claude. ## Data residency for gpt-oss-20b On inference.eu the open weights of gpt-oss-20b run on infrastructure in the European Union. Prompts and responses are processed in EU data centres, nothing is sent to OpenAI, and nothing is used to train any model. That is the main difference from calling OpenAI's hosted models. Every customer receives a data processing agreement under Article 28 of the GDPR and our list of subprocessors, so your records of processing can name every party involved. Billing is in euros. The Apache 2.0 licence permits commercial use without user limits; OpenAI adds a short usage policy that asks you to comply with applicable law. ## gpt-oss-20b Maker OpenAI Made in United States Type Reasoning Parameters 21B Active parameters 3.6B Context window 131,072 tokens Licence [Apache 2.0](https://huggingface.co/openai/gpt-oss-20b/blob/main/LICENSE) Released August 2025 Model ID gpt-oss-20b Processing EU data centres Training on your data Never Source [huggingface.co](https://huggingface.co/openai/gpt-oss-20b) gpt-oss-20b · Python ``` import os from openai import OpenAI client = OpenAI(base_url="https://api.inference.eu/v1", api_key=os.environ["INFERENCE_EU_API_KEY"]) response = client.chat.completions.create( model="gpt-oss-20b", messages=[{"role": "user", "content": "Summarise this contract in five bullet points."}], ) print(response.choices[0].message.content) ``` ## What teams use it for - 1 ### Fast support answers Answer routine customer questions with a short reasoning step, at a size that keeps response times low. - 2 ### Tool routing Decide which internal API or workflow should handle a request and fill in the parameters as structured JSON. - 3 ### Data extraction with checks Pull fields from emails and documents and reason briefly about whether the values are consistent. ## gpt-oss-20b: common questions The questions buyers, developers and legal teams ask us most. Something missing? Ask us directly. Compliance depends on how you process data. On inference.eu, gpt-oss-20b requests are processed in EU data centres, not used for training and covered by a data processing agreement under Article 28 GDPR with a published subprocessor list. Prompts are not sent to OpenAI or transferred to the United States. gpt-oss-20b supports 128K tokens, or 131,072. Reasoning tokens count towards the output, so leave room for them when you use medium or high reasoning effort. For most chats and documents the window is more than enough. Yes. The model is released under the Apache 2.0 licence, which allows commercial use, modification and redistribution without user limits. OpenAI's accompanying gpt-oss usage policy mainly asks you to comply with applicable law. gpt-oss-20b has 21 billion parameters with 3.6 billion active; gpt-oss-120b has 117 billion with 5.1 billion active. The 120b model gives better answers on difficult reasoning and coding. The 20b model is faster and lighter, which suits simple and high-volume tasks. Set \`reasoning\_effort\` to \`low\`, \`medium\` or \`high\` in each request. Low suits general dialogue, medium balances speed and detail, and high produces deeper analysis at the cost of more output tokens and a slower response. No. gpt-oss-20b is text-only. For images and screenshots, use a multimodal model such as Gemma 4 26B A4B or Qwen3.6 35B A3B, which are similarly light per token, or DeepSeek-OCR 2 for scanned documents. It comes from OpenAI, but it is a separate open-weight model, not the model behind ChatGPT. Anyone can download and run it. On inference.eu it runs on our EU infrastructure, independent of OpenAI's own services. Both are compact Apache 2.0 reasoning models. gpt-oss-20b is a mixture-of-experts model with fewer active parameters and adjustable effort. Magistral Small is a dense 24B model from Mistral AI in France that also reads images. Test both on your workload. In your OpenAI client, set the base URL to the inference.eu API, use your inference.eu key and change the model to \`gpt-oss-20b\`. Tools, structured outputs and streaming keep working with the same code. No. OpenAI published the weights but takes no part in serving your requests on inference.eu. Your prompts are processed in EU data centres, used only for your response and never used to train any model. [Ask a question](https://inference.eu/contact/) ## Related [All models](https://inference.eu/models/#directory) - [gpt-oss-120bRun OpenAI's open-weight gpt-oss-120b reasoning model on inference.eu: the OpenAI API format, with requests processed in EU data centres, billed in euros.Reasoning](https://inference.eu/models/gpt-oss-120b/) - [Magistral Small 1.2Run Magistral Small 1.2, Mistral AI's open 24B reasoning model, on inference.eu through an OpenAI-compatible API with processing in EU data centres.Reasoning](https://inference.eu/models/magistral-small/) - [Gemma 4 26B A4BRun Google's Gemma 4 26B A4B on inference.eu: a fast mixture-of-experts model with image input and 256K context, processed in EU data centres.Vision](https://inference.eu/models/gemma-4-26b-a4b/) - [Qwen3.6 35B A3BRun Qwen3.6 35B A3B on inference.eu: a fast mixture-of-experts model with 3B active parameters, image input and 256K context, processed in the EU.Code](https://inference.eu/models/qwen3-6-35b-a3b/) ## Move your AI workloads to EU infrastructure Tell us which models you use and roughly how much traffic you have. We reply within one working day with pricing and next steps. - Reply within one working day - Pricing for your volume - DPA ready to sign