Mistral announces Voxtral 3

Mistral has announced Voxtral 3, a set of speech models paired with a voice agent SDK. The idea is to handle the whole spoken interaction in one real-time process: listen, work out what the person needs, use connected tools, then respond. In Mistral’s example, an agent reads a referral letter in French, explains it in English, and books an appointment.

For a business, the practical point is that a voice agent could do more than answer common questions. Mistral says agents can retrieve information, interact with business systems, and complete multi-step tasks during phone or web conversations. The models cover more than 50 languages for transcription, while the text-to-speech model supports more than 35 languages and 50 dialects. Organizations can use Mistral’s EU-hosted API or deploy the system on their own infrastructure, according to the Voxtral 3 announcement.

That makes this a development to track if customers or staff already handle work by phone, especially where a conversation needs to lead to an action in another system. The hard part is making those steps work accurately while the conversation is happening, so the announcement is not proof that every workflow is ready to automate. It is the kind of capability we think teams should consider when evaluating an AI command centre.