Mistral ships Voxtral 3: speech models plus a LiveKit-based voice agent SDK

Mistral released Voxtral 3, a family of transcription and TTS models with a voice agent SDK, available via EU, US and global API endpoints.

Mistral released Voxtral 3 on October 20, 2026. It is a family of speech-to-text and text-to-speech models plus a voice agent SDK, built on LiveKit, that runs listening, reasoning, tool calls and speech as one streaming pipeline. If you build phone or web voice agents, you can get the whole stack from one vendor, with the option to keep audio processing in Europe or on your own servers.

Mistral's price list:

According to Mistral, agents can place and answer calls over SIP, run over WebRTC, and swap the LLM or voice without a rewrite. The API runs on European, US and global endpoints. On-prem deployment goes through Mistral's sales team. THE DECODER confirms the model lineup, the 50+ languages and the LiveKit base.

Treat the performance claims with care. CellCog notes that every benchmark in the post is Mistral's own, that the speech evals were judged by gemini-3.8-flash, and that Mistral's own realtime chart shows Gemini 3.5 Realtime Transcribe ahead at every latency setting. CellCog also says the post doesn't mention publishing weights. The "under one second per turn" figure is Mistral's. Tech Times points out that a SIP-to-WebRTC bridge adds 50–150 ms that Mistral's figures don't include.

Alternatives include the vendor-neutral, open-source Pipecat framework and ElevenLabs' Eleven Realtime speech-to-speech models, which run through an OpenAI-compatible Realtime API.

Sources