Mistral released Voxtral 3 on Tuesday, a family of models aimed at voice agents that need to listen, reason, call tools and answer inside a live conversation.
The bundle holds two models and a toolkit. Voxtral Small Transcribe 3 handles speech recognition across more than 50 languages. Voxtral Mini TTS 3 generates speech in over 35 languages and 50 dialects, with emotional adaptation. A voice agent SDK built on LiveKit binds them into one streaming pipeline.
Tool calling sits inside that loop. An agent can reach external systems mid-conversation to book an appointment or check an invoice without breaking the audio stream.
Deployment is the differentiator
Mistral says the stack runs on customer infrastructure or through EU-hosted APIs, which keeps audio and transcripts under the buyer’s control. For European enterprises weighing voice agents against data-residency obligations, that is the argument that matters.
Transcription starts at $0.0035 per minute. Volumes at that level change what a support line costs to operate, and the price puts Mistral alongside the cheapest hosted options rather than above them.
Voice has become the most contested interface in AI. Google shipped two live voice models in recent weeks, and every major lab now offers some combination of recognition, synthesis and orchestration.
What Mistral is selling is the assembled stack. Most competitors leave integration to the customer, and stitching a real-time pipeline together reliably is where voice projects usually stall.
