Voxtral 3
Points and comments are a snapshot, not live.
Mistral launches Voxtral 3, a voice AI family with STT, TTS, and an SDK for real-time agent pipelines.
Mistral's Voxtral 3 includes Voxtral Small Transcribe 3 (state-of-the-art in 50+ languages), Voxtral Mini TTS 3 (35+ languages, $13/1M chars), and a voice agent SDK combining listening, reasoning, tool calling, and speaking in a streaming pipeline. The SDK, built on LiveKit, targets contact centers and virtual assistants, supporting real-time conversation with sub-second turn completion. Pro Voice adaptation reproduces a speaker's style from 30 minutes of audio. Models are available via Mistral's EU-hosted API or for self-deployment on private infrastructure.
What commenters are saying
Commenters widely note the models are closed-source, contrasting with earlier Mistral releases. Several recommend alternatives for local deployment: Parakeet and OpenMOSS for STT, Kokoro and Gradium's new 31M-param TTS for speech generation. Some see the closed approach as necessary given EU regulation and threat models around voice data. A few questioning the demo's realism about real-time latency.