Mistral launches Voxtral 3: real-time speech models and a voice agent SDK

What happened

Mistral released Voxtral 3 — a new generation of speech models and a voice agent SDK available via the Mistral API and as open weights on Hugging Face.

Context and impact

Voxtral 3 competes with OpenAI Realtime API and ElevenLabs. The key differentiator is self-hosting (open weights), data control, and performance per dollar. Target use cases: contact centers, hands-free copilots, virtual assistants.

Details

• Transcription in 50+ languages with speaker diarization and domain biasing without fine-tuning
• Expressive TTS across 35+ languages — emotion, pacing, and emphasis adapt to the conversation
• Voice adaptation: cloning from 30+ minutes of audio; Pro Voice captures full speaking style
• Voice Agent SDK (LiveKit, SIP/WebRTC): each turn completes in under one second
• Voxtral Small Transcribe 3: state-of-the-art on FLEURS multilingual benchmark, ahead of Whisper large-v3
• Evaluation data generated on 2026-10-20