Mistral introduces Voxtral 3 voice models and agent SDK

Mistral presents Voxtral 3 as a family of speech-to-text and text-to-speech models alongside a voice agent SDK.[2] It says the SDK is built on LiveKit and streams transcription, reasoning, tool calls and speech as one pipeline.[1] Mistral’s stated design assigns distinct roles to the models: a transcription model listens, a reasoning model thinks and calls tools, and a speech model replies.[1] For teams assessing voice-agent architecture, the announcement describes an integrated flow rather than a single model handling every stage. That description is Mistral’s account of the SDK, not an independent performance assessment.

Mistral says Voxtral agents are intended for contact centers, service desks, virtual assistants and hands-free copilots.[2] It says agents can retrieve information, interact with business systems and complete tasks during live phone or web conversations.[2] Those use cases make the tool-calling part of the pipeline particularly relevant to teams considering whether an agent should do more than transcribe or answer questions. Mistral describes the listening, thinking, tool-calling and speaking stages as a single real-time streaming pipeline.[2] The announcement does not establish how well the system performs in any one setting.

Mistral says Voxtral Small Transcribe 3 includes diarization, context-biasing and end-of-turn detection, and supports 24 official EU languages plus languages with fewer resources in Asia and India.[2] Its Mini transcription model has 4 billion parameters and costs $0.0035 per minute, according to Mistral.[2] Voxtral Mini TTS 3 supports more than 35 languages and 50 dialects, with speech that Mistral says adapts emotion, pace and emphasis to context.[2] Mistral also says Pro Voice adaptation can reproduce a speaker’s character, emotional range and style across languages using 30 minutes of audio.[2] It lists Voxtral TTS at $13 per million characters.[2]

Mistral says Voxtral 3 can be accessed through its EU-hosted API or self-deployed on an organization’s own infrastructure.[2] It says those choices let organizations control where audio is processed, how data is handled and whether conversations are retained.[2] MarkTechPost reports that all four Voxtral models were in public preview on the Mistral API at release.[3] It also reports that the SDK handles each turn in under one second and supports phone calls over SIP, plus browser or app sessions over WebRTC.[3] MarkTechPost gives voice-agent API pricing as $0.04 per minute with Mistral Small 4, or $0.08 with another reasoning model.[3]