Use LangChain with open-source models hosted in the EU

LangChain is an open-source framework in Python and JavaScript for building applications on language models. It provides a common interface for chat models, prompts, tools, retrievers and output parsers, and its create_agent helper and LangGraph library run tool-calling agents and stateful workflows.

Every LangChain component that needs a chat model accepts any object implementing that common interface. The ChatOpenAI class from langchain-openai talks to any endpoint in the OpenAI format when you set base_url, so inference.eu slots into existing chains and agents without other code changes.

Pass base_url (the inference.eu base URL ending in /v1), api_key and model to ChatOpenAI. In LangChain.js, the URL goes under configuration: { baseURL }. LangChain also reads OPENAI_API_BASE and OPENAI_BASE_URL when no explicit URL is passed, but explicit arguments are clearer when a project uses several providers.

With a custom base URL, LangChain does not turn on streaming usage by default, so set stream_usage=True to receive token counts. You can also build the model with init_chat_model("openai:<model id>", base_url=..., api_key=...) if your code selects models from configuration.

For retrieval-augmented generation, summarisation and extraction, gpt-oss-120b is a quick and capable default. For agents built with create_agent or LangGraph that call several tools in sequence, qwen3-coder-480b-a35b-instruct, glm-5.3 and kimi-k3 produce dependable tool calls.

glm-5.3 and kimi-k3 also suit long-context work over large documents and complex multi-step planning. Use with_structured_output with a Pydantic model for extraction; LangChain validates the result. Because models are plain objects, you can give each node in a graph its own model, using a fast one for routing and a stronger one for reasoning.

A 401 error means the key is missing or LangChain picked up an old OPENAI_API_KEY; pass api_key explicitly. A 404 usually means base_url has an extra path, or the model id does not exist at inference.eu.

ChatOpenAI follows the official OpenAI specification only, so non-standard fields such as reasoning_content from reasoning models are not kept in the message. Using built-in tools or parameters that only exist in OpenAI's Responses API switches ChatOpenAI to that API; keep use_responses_api=False and stick to Chat Completions features. If tool calls arrive as text, choose a model with tool calling support.

With base_url pointed at inference.eu, every prompt, retrieved document and tool result your chains send to the model is processed in EU data centres and never used for training. We sign a data processing agreement under Article 28 GDPR, publish our subprocessor list and bill in euros.

LangChain itself only calls the services you configure. Embeddings, vector stores and tracing are separate: if you enable LangSmith tracing, prompts and outputs are sent to LangSmith under its own terms. Choose EU-hosted options for those components, or keep tracing off, so that the whole pipeline stays in Europe.