Mistral launches Voxtral 3, a full real-time voice-agent stack

Mistral has released Voxtral 3, an end-to-end stack for building real-time voice agents. It combines Voxtral Small Transcribe 3 for speech recognition, Voxtral Mini TTS 3 for speech synthesis, and an agent SDK built on LiveKit, the open-source real-time media framework. Together they form one streaming pipeline that handles listening, reasoning, tool calls and spoken responses.

The numbers are competitive. Voxtral Small Transcribe 3 reports a 4.1% word error rate on the FLEURS multilingual benchmark, ahead of OpenAI's Whisper, at $0.004-0.008 per minute. Building on LiveKit gives developers a familiar transport layer for WebRTC audio, so the stack fits into telephony and app integrations without a proprietary client.

Competitively, Voxtral 3 is aimed at OpenAI's Realtime API and at proprietary voice platforms. Voice-agent builders in the community are drawn to it as an open, end-to-end alternative. The launch came alongside Timestral, Mistral's time-series family, and follows a €3B Series D at a €21B valuation that has given the company room to broaden beyond text models.

The caveats: real-time voice quality depends heavily on end-to-end latency and interruption handling, which benchmarks like FLEURS don't capture. Developers will want latency numbers under load and language coverage for TTS. Watch for production deployments in European customer-service settings, where Mistral's sovereignty pitch is strongest.