Mistral AI launches Voxtral 3, a new family of speech models for versatile voice agents

Mistral AI has released Voxtral 3, a family of speech models bundled with a voice agent SDK rather than a single flagship model. According to the company's announcement, the release comprises three transcription models — Voxtral Small Transcribe 3, Voxtral Mini Transcribe 3, and Voxtral Mini Transcribe Realtime 3 — plus Voxtral Speech 3 for synthesis, alongside an SDK built on LiveKit that streams speech recognition, reasoning, and speech generation as one continuous pipeline. Mistral says each stage runs independently, so developers can swap the reasoning model, voice, or language without rebuilding the agent, and that turns complete in under a second with support for interruption. The models are available today through Mistral's API on European, US, and global endpoints, and for self-deployment on customer infrastructure, a sovereignty pitch Mistral frames as central to the launch, aimed at contact centers, service desks, and hands-free copilots.

Performance claims come largely from Mistral's own evaluations. The announcement cites an internal benchmark showing Voxtral Speech 3 leading competitors on expressiveness per dollar, and third-party tracker Artificial Analysis placed Voxtral Small Transcribe 3 third on its speech-to-text leaderboard. Price puts batch transcription at $0.0035 per minute for Mini, $0.004 for Small, and $0.008 for the real-time model, with speech synthesis at $13 per million characters.

Language coverage is where accounts diverge slightly: Mistral's announcement says transcription spans 50+ languages, while it seems the offering really spans across 35+ languages and 50+ dialects. The louder reaction, though, concerned licensing. Voxtral 3 is closed-weight, a departure from the previous Voxtral generation, prompting threads recommending open-source alternatives, with users also pressing on barge-in latency and noisy-audio accuracy that the announcement leaves unaddressed. Mistral's post doesn't indicate whether weights will be opened later, as the models currently sit in public preview.