Mistral Voxtral 3, a model family and voice agent SDK

Introduction

A voice agent chains three building blocks : transcription, reasoning and speech synthesis. On October 20, 2026, Mistral AI launched Voxtral 3, a family of four voice models in public preview, along with a voice agent SDK. Mistral now provides all three building blocks, with transcription expanded to 50 languages.

Check the Voxtral Transcribe 2 article HERE.

Transcription : Small, Mini and Realtime

Batch transcription relies on two models. Voxtral Small Transcribe 3 (voxtral-small-transcribe-3-0) is the most accurate in the family ; its size hasn’t been published. Voxtral Mini Transcribe 3 (voxtral-mini-transcribe-3-0), with 4 billion parameters, is built for speed. Both cover 50 languages, including the 24 official languages of the European Union, with diarization and context biasing, which accepts a list of key terms or free text.

Voxtral Mini Transcribe Realtime 3 (voxtral-mini-transcribe-realtime-3-0) is designed for agents : Mistral claims latency under 500 ms in more than 35 languages. A single WebSocket session brings together end-of-turn detection, context biasing and diarization.

Text-to-speech with Voxtral TTS 3

Voxtral TTS 3 (voxtral-mini-tts-3-0) succeeds Voxtral TTS, released in March 2026. According to the announcement, it covers more than 35 languages and more than 50 dialects.

Features include :

• Instructions : Prompt-based instruction following and contextual synthesis
• Voice control : Emotion, pacing and emphasis, with 34 emotion tags and 2 pacing tags (slow and fast)
• Streaming : About 90 ms to first audio
• Formats : Five output formats, including 24 kHz WAV
• Voice cloning : Zero-shot from a short recording, or professional cloning (Pro Voice), for which the documentation recommends 30 minutes to 3 hours of single-speaker audio

A cloned voice isn’t limited to the language of its recording. It’s kept for 30 days by default, adjustable from 1 to 365 days through the API, and the reference clip is sent as base64, never as a URL.

An SDK for voice agents

The voice agent SDK relies on LiveKit for real-time audio transport. It exposes Voice Agents, a bidirectional voice pipeline managed by Mistral. The agent answers phone calls over SIP and runs in browsers and apps over WebRTC. No stage waits for the previous one, which, according to Mistral, keeps each conversational turn under one second.

With the browser SDK, the server creates the session : the API key never reaches the browser. The model used at each stage can be chosen per agent or per session.

What’s changed since Voxtral Transcribe 2

Compared with the previous generation :

• Languages : Transcription goes from 13 to 50 languages
• Batch transcription : From 0.003$ per minute for Voxtral Mini Transcribe V2 to 0.0035$ for Mini Transcribe 3
• Real time : From 0.006$ to 0.008$ per minute
• Text-to-speech : From 16$ / 1M characters for Voxtral TTS to 13$ / 1M characters
• Open weights : Voxtral Realtime (Apache 2.0) and Voxtral TTS (CC BY-NC 4.0) were released as open weights ; the Voxtral 3 announcement mentions neither open weights nor a license

The Voxtral 2 documentation is now marked Deprecated, with a migration guide but no retirement date.

Pricing and availability

Voxtral 3 is available in public preview through the Mistral API, billed by usage, and can be tried in the Mistral Studio playgrounds. Prices, in USD :

Voice agents are billed per minute of conversation : 0.04$ with Mistral Small 4 for reasoning, 0.08$ with a larger model that Mistral doesn’t name. Mistral positions its text-to-speech pricing as lower than that of OpenAI TTS-1, Deepgram, ElevenLabs Flash and Gemini Flash.

Conclusion

Voxtral 3 makes Mistral an end-to-end provider for voice agents, with pay-as-you-go pricing that’s easy to estimate. For a Teams and voice admin, it isn’t a contact center solution yet : telephony remains reserved for enterprise customers, inbound only and with no transfer to a person. Today, Voxtral 3 is mainly a prototyping tool ; what to watch is telephony opening up and any open-weights announcement, since open weights are what made Voxtral 2 attractive for self-hosting.

Sources

Mistral AI - Voxtral 3

Mistral AI Docs - Telephony (SIP)

Try Le Chat from Mistral AI

Did you enjoy this post ? If you have any questions, comments or suggestions, please feel free to send me a message from the contact form.

Don’t forget to follow us and share this post.