Mistral releases Voxtral 3 speech models and a voice agent SDK

Mistral released the Voxtral 3 family on 20 October 2026, combining accurate multilingual transcription, expressive text-to-speech and a real-time voice agent SDK that handles listening, reasoning, tool calling and speaking in one pipeline [1]. The models run on Mistral's EU-hosted API or on customers' own infrastructure [1].

What transcription models did Mistral ship?

Mistral released three Voxtral 3 transcription models on 20 October 2026, supporting up to 50 languages. They are Voxtral Small Transcribe 3, Voxtral Mini Transcribe 3 and Voxtral Mini Transcribe Realtime 3 [1]. The Mini model has 4B parameters and costs $0.0035 per minute while the realtime variant targets sub-500 ms latency in 35+ languages with diarization and context biasing [1].

Voxtral Small Transcribe 3 claims state-of-the-art results on diarization and multilingual transcription across 50 languages including 24 official EU languages, outperforming Whisper, Scribe v2 and Gemini 3.5 Transcribe according to Mistral's benchmarks [1]. Context biasing lets users supply key terms without fine-tuning [1].

How does the new TTS and voice cloning work?

Mistral says Voxtral Mini TTS 3 supports more than 35 languages and 50 dialects at $13 per million characters. Mistral also calls the model Voxtral Speech 3, and says its voices adjust emotion, pace and emphasis to the conversation [1]. For professional use, Mistral says voice adaptation can reproduce a speaker's style from 30 minutes of audio [1].

What does the voice agent SDK enable?

Mistral's new voice agent SDK is built on LiveKit and completes each turn in under one second, Mistral says. It runs speech in, reasoning and speech out as one streaming pipeline, and the agent stops talking when the caller cuts in [1]. Agents can answer and place calls over SIP and run in browsers and apps over WebRTC [1].

Where can developers run Voxtral 3?

Voxtral 3 is available through the Mistral API with European, US and global endpoints, and can be self-deployed [1][2]. Mistral says organisations control where audio is processed and whether any audio is retained [1]. Mistral's benchmark charts, which claim wins over Whisper, Scribe v2 and Gemini 3.5 Transcribe, are its own evaluations [1].

Our read

Opinion: Voxtral 3 closes the gap between speech interfaces and reliable agent workflows, but the real test will be whether the claimed accuracy and latency hold once independent benchmarks replace Mistral's internal numbers.

What we could not verify

• Independent confirmation of the Word Error Rate and diarization numbers shown in Mistral's charts.
• The exact latency distribution of the realtime model under production load.
• Whether Voxtral 3 weights are published openly; Mistral's post describes API access and self-deployment.
• Any third-party evaluation of the 30-minute Pro Voice adaptation quality.