Voxtral 3

Summary

Mistral released Voxtral 3, a family of speech models for voice agents: two batch transcription models, a realtime transcription model, a text-to-speech model, and a Voice Agent SDK built on LiveKit. The post has no explicit date. Mistral's news index lists it on 2026-10-20, and its benchmarks were generated that day. The post does not say whether the weights are open, but it offers self-hosting.

Key points

• Models (documentation slugs, likely the API ids): voxtral-small-transcribe-3-0, voxtral-mini-transcribe-3-0 (4B), voxtral-mini-transcribe-realtime-3-0 (under 500 ms latency), and voxtral-mini-tts-3-0.
• Transcription: diarization, keyword biasing, and end-of-turn detection. Batch covers 50+ languages and realtime 35+ (the post also says 40+ in one place).
• TTS: emotion, pacing, and emphasis control, and voice cloning from about 30 minutes of audio. It covers 35+ languages and 50+ dialects.
• Voice Agent SDK: chains transcription, an LLM, and TTS with function calling, web search, and retrieval. It supports SIP telephony and WebRTC.
• Prices (2026-10-20): batch transcription $0.0035/min (Mini) and $0.004/min (Small); realtime $0.008/min; TTS $13 per 1M characters; voice agent API $0.04/min with Mistral Small 4.
• Benchmarks (self-reported): about 4.1% WER on FLEURS for Small, against about 16% for Whisper large-v3. Mistral says it beats Scribe v2 and Gemini 3.5 Transcribe.
• Availability: Mistral API (EU, US, and global endpoints), Studio playgrounds, and self-hosting. Contact Mistral for on-premises deployment.

Why it matters

It is a full European voice-agent stack at low per-minute prices, and its transcription is a strong alternative to Whisper. It is unclear whether the weights are open as Voxtral 1 (Apache 2.0) was; this is noted as an open question.