Run Voxtral Small in the EU for transcription and audio understanding

Voxtral Small 1.0, published as Voxtral-Small-24B-2507 in July 2025, is a 24-billion-parameter model from Mistral AI that adds audio input to the Mistral Small 3 language model. It can run in a dedicated transcription mode, or take audio and text together and answer questions, summarise or follow spoken instructions.

It has a 32K-token context, which Mistral says covers up to about 30 minutes of audio for transcription and 40 minutes for understanding. It detects the spoken language automatically and is strongest in English, Spanish, French, Portuguese, Hindi, German, Dutch and Italian. The weights are released under Apache 2.0.

Voxtral's advantage over a pure transcription model is that it understands what it hears. You can ask it for a summary of a call, the action items from a meeting or the answer to a question asked in a voice note, in a single request and without a separate language model. It can also trigger function calls directly from spoken requests.

It is a large model for transcription alone, so for high-volume, plain speech to text a smaller dedicated model such as Whisper Large V3 Turbo can be more efficient. Recordings longer than about 30 to 40 minutes need to be split. Mistral notes that system prompts are not yet supported, so put instructions in the user message.

inference.eu uses the OpenAI request format. For plain transcription, send the audio file to the OpenAI-style transcription endpoint with the model set to voxtral-small-24b, exactly as you would with OpenAI's speech models after changing the base URL and key.

To ask questions about a recording, send the audio as an input_audio content part in a chat message together with your text question. Several audio clips and several turns in one conversation are supported. Mistral recommends temperature 0 for transcription and around 0.2 for understanding tasks. Tool definitions work as with any chat model.

Voice recordings often contain names, health details or customer data, and a voice can itself be personal data. With Voxtral Small on inference.eu, audio and transcripts are processed in data centres in the European Union, used only to produce your result and never used to train any model.

Every customer receives a data processing agreement under Article 28 of the GDPR and our list of subprocessors, so you can document how recordings are handled. Billing is in euros. Voxtral Small comes from a French lab and is licensed under Apache 2.0, which allows commercial use without user limits. You remain responsible for informing callers that they are recorded.