Mistral AI has launched Voxtral 3, a comprehensive family of language models designed specifically for real-time voice agents. The new stack integrates a speech-to-text model called Voxtral Small Transcribe 3, a text-to-speech model named Voxtral Mini TTS 3, and a dedicated voice agent SDK built on LiveKit to combine listening, thinking, tool use, and speaking into a unified streaming pipeline.
The transcription capabilities cover more than 50 languages, while the text-to-speech output supports over 35 languages and 50 dialects featuring emotional adaptation. Furthermore, these voice agents are capable of interacting with external systems during live conversations to perform practical tasks such as booking appointments or checking invoices.
Organizations can deploy the system either on their own infrastructure or utilize EU-hosted APIs to ensure data remains under full control, with transcription pricing starting at $0.0035 per minute. This release matters because it provides enterprises with a powerful, flexible open infrastructure stack for building real-time voice assistants with robust multilingual support.