Use Mistral Vibe with open-source models hosted in the EU

Mistral Vibe is an open-source command-line coding agent from Mistral AI. It reads and edits files, runs shell commands, searches your repository and can delegate work to subagents, with permissions you control per tool. It also supports MCP servers, custom agents and hooks.

By default Vibe talks to Mistral's API, but providers and models are plain entries in config.toml. A provider with the generic backend sends standard OpenAI Chat Completions requests to whatever api_base you set, which is how you point Vibe at inference.eu.

Vibe reads ~/.vibe/config.toml. The [[providers]] block gives the provider a name, sets api_base to the inference.eu base URL ending in /v1, and names the environment variable holding your key in api_key_env_var. api_style = "openai" and backend = "generic" tell Vibe to use the standard OpenAI wire format.

Each [[models]] block links an inference.eu model id (name) to the provider and gives it an alias. active_model refers to that alias. Use --model <alias> to try another model for a single run without changing the default, or switch inside a session from the model picker.

Vibe works in a loop of reading, editing and running commands, so it needs a model that calls tools reliably and keeps track of a growing context. qwen3-coder-480b-a35b-instruct is a strong default: it is trained for agentic coding and produces precise edits.

For architecture questions, larger refactors or debugging that needs more reasoning, add glm-5.3 or kimi-k3 as a second alias and switch when needed. gpt-oss-120b is quick and suits subagents or short tasks. The default temperature for custom models is low, which suits code; raise it per model only if you want more varied output.

A 401 error means Vibe sent no key or the wrong one. Check that INFERENCE_EU_API_KEY is exported in the shell that starts Vibe and that api_key_env_var names exactly that variable. A 404 usually comes from an api_base with an extra path.

An "unknown alias" error means active_model or --model does not match any alias. "Model not found" from the API means name differs from the inference.eu id. If Vibe reports incomplete streams and retries, check you are on a current version; emits_finish_reason = false on the provider is only needed for servers that end streams without a finish reason.

With inference.eu as the provider, the files, diffs and command output Vibe sends to the model are 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.

Vibe can also send usage telemetry and crash reports to Mistral, separately from model requests. Set enable_telemetry = false in config.toml to turn this off. Mistral AI is a French company, but telemetry is governed by its own privacy policy, so disabling it keeps your sessions limited to traffic between Vibe and inference.eu.