Overview
Mistral Voxtral 3 is a speech model family released by Mistral AI on October 20, 2026, with the official page titled Voxtral 3: Voice Agent SDK & Speech Models in 50+ Languages. The family includes both speech transcription and text-to-speech models, along with Voice Agents API capabilities, covering accurate transcription in more than 50 languages. Voxtral Small Transcribe 3 ranks 3rd on the Artificial Analysis speech-to-text leaderboard, and the TTS model enters the top 12. Users can call it through Mistral's EU-hosted API or choose self-hosted deployment. As a new-generation family following the original open-source Voxtral speech model, Voxtral 3 targets development teams and enterprise scenarios that need multilingual speech recognition, speech synthesis, and voice agent construction.
Key Features
- Dual-mode speech capabilities: The family provides both speech transcription and text-to-speech models, covering the two core speech tasks of recognition and synthesis.
- More than 50 languages: Supports accurate transcription in more than 50 languages, making it easy to build speech applications for multilingual user groups.
- Voice Agents API: Comes with Voice Agents API capabilities, providing interface support for building voice agent applications.
- EU-hosted calls: Can be called through Mistral's EU-hosted API, meeting scenarios with requirements on data residency regions.
- Self-hosted deployment: In addition to the hosted API, self-hosted deployment is also supported, making it easy to run speech models in your own environment.
- Leaderboard performance: Voxtral Small Transcribe 3 ranks 3rd on the Artificial Analysis speech-to-text leaderboard, and the TTS model enters the top 12.
Use Cases
- Input multilingual meeting recordings into Voxtral Small Transcribe 3 for transcription to generate searchable text records.
- Use the text-to-speech model to synthesize product documents or notification content into speech for multilingual podcasts or prompt tones.
- Build voice agents through the Voice Agents API to complete voice interaction flows in customer service or assistant scenarios.
- Enterprises with data residency requirements can choose the EU-hosted API to call speech transcription and synthesis capabilities within compliant regions.
- Teams with their own infrastructure can self-host and deploy Voxtral 3 models to process speech data in their own environment.
- For markets covering more than 50 languages, use the same family of models to complete multilingual speech transcription and speech output.
Pros
- The official page clearly states that the family covers both speech transcription and text-to-speech models, with a clear capability scope.
- Voxtral Small Transcribe 3 ranks 3rd on the Artificial Analysis speech-to-text leaderboard, verifiable through a public leaderboard.
- The TTS model enters the top 12 of the relevant Artificial Analysis leaderboard, providing a third-party ranking reference for synthesis capability.
- Supports accurate transcription in more than 50 languages, suitable for direct use in multilingual business scenarios.
- Provides two access methods, the EU-hosted API and self-hosted deployment, offering flexible deployment choices.
- As a new-generation family following the original open-source Voxtral speech model, it continues Mistral's investment in the speech direction.
Pricing
The specific pricing and quotas for Voxtral 3 are subject to the official website; the official page does not list price figures in the given information. Billing methods differ by the API calls and deployment methods provided by Mistral: usage through the EU-hosted API is billed per call, while self-hosted deployment involves bearing infrastructure costs yourself. Access channels include the Mistral EU-hosted API and self-hosted deployment; it is recommended to refer to the latest billing rules and quotas published on the official website.
Summary
Mistral Voxtral 3 is a speech model family released on October 20, 2026, including speech transcription and text-to-speech models, along with the Voice Agents API, supporting more than 50 languages. Voxtral Small Transcribe 3 ranks 3rd on the Artificial Analysis speech-to-text leaderboard, and the TTS model enters the top 12. Users can call it through the EU-hosted API or deploy it self-hosted; specific pricing and quotas are subject to the official website.
Version History
- Voxtral 3 launches: speech-to-text and text-to-speech family covering 50+ languages (2026-10-20): Mistral AI released the Voxtral 3 speech model family on October 20, 2026. The family covers speech-to-text and text-to-speech models alongside a Voice Agents API. Voxtral Small Transcribe 3 ranks third on the Artificial Analysis speech-to-text leaderboard, with TTS models in the top 12. It supports transcription in more than 50 languages and is available through Mistral's EU-hosted API or self-hosted deployment.