Mistral AI has introduced the Voxtral3 voice model series, along with its complementary voice agent SDK. This series encompasses transcription and speech synthesis models, in addition to a LiveKit-based voice agent SDK. It supports both API calls and self-hosted deployments, enabling high-performance, real-time streaming voice interactions across a variety of scenarios. The transcription model boasts support for over 50 languages, with several error rate metrics that are on par with Microsoft's equivalent offerings and outperform most mainstream competitors. The version with 4 billion parameters achieves real-time latency of under 500 milliseconds and offers transparent pricing. The speech synthesis model supports more than 35 languages and over 50 dialects, providing features such as tone adaptation and voice cloning. The complementary SDK integrates full-process capabilities, allowing dialogue turn response times to be controlled within 1 second and supporting automatic agent shutdown during interruptions, thereby facilitating the widespread adoption of voice AI agents.