Voxtral (Mistral AI)

<p>Voxtral is Mistral AI's family of speech models. It turns spoken audio into text and can also answer questions about, or summarise, what was said. Some versions are released as open weights under the Apache 2.0 licence, which means developers can download them and run them on their own servers at no licence cost.</p> <h2>What it does</h2><p>Mistral launched the first Voxtral models in July 2025 in two sizes: Voxtral Small (24 billion parameters) for production use and Voxtral Mini (3 billion parameters) for local and edge deployment. Both could transcribe audio and work with recordings directly, so a single model could summarise a meeting or pull out answers without a separate language model.</p> <p>In February 2026 Mistral released <a href="https://mistral.ai/news/voxtral-transcribe-2" rel="noopener">Voxtral Transcribe 2</a>, which focuses on speech-to-text:</p> <ul><li><strong>Voxtral Mini Transcribe V2</strong> handles recorded files up to three hours long. It adds speaker diarisation (who said what), word-level timestamps and "context biasing", where you supply up to 100 names or terms so it spells them correctly.</li><li><strong>Voxtral Realtime</strong> transcribes live audio with a delay that can be set below 200 milliseconds. It is the open-weight model in this release, at about 4 billion parameters, and small enough to run on edge devices.</li></ul><h2>Languages</h2><p>Voxtral Transcribe 2 supports 13 languages: English, Chinese, Hindi, Spanish, Arabic, French, Portuguese, Russian, German, Japanese, Korean, Italian and Dutch. It can follow speakers who switch between languages in the same conversation.</p> <p>No African languages other than Arabic are on the list. Morocco's Ministry of Digital Transition has used Mistral's Voxtral technology as the base for an open-source Darija speech recognition model, which shows the models can be adapted, but that work is separate from Mistral's own product.</p> <h2>Pricing</h2><ul><li><strong>Open-weight models:</strong> free to download from Hugging Face under Apache 2.0. You pay only for your own hosting.</li><li><strong>Voxtral Mini Transcribe V2 (API):</strong> $0.003 per minute of audio.</li><li><strong>Voxtral Realtime (API):</strong> $0.006 per minute.</li></ul><p>At those rates, transcribing a one-hour interview through the API costs about 18 US cents. Prices are from Mistral's announcement and may change.</p> <h2>How to access it</h2><ul><li>Through Mistral's API, at the <code>/v1/audio/transcriptions</code> endpoint</li><li>In Mistral Studio's audio playground, which accepts MP3, WAV, M4A, FLAC and OGG files up to 1 GB</li><li>In Le Chat, Mistral's chat assistant</li><li>As open weights on Hugging Face, for self-hosting</li></ul><h2>Who it suits</h2><p>Voxtral is a good fit for developers building transcription into an app, call-centre or media teams with large volumes of recorded audio, and organisations that need audio data to stay on their own servers. Teams working mainly in African languages other than Arabic will need a different tool, or will have to fine-tune the open model themselves.</p> <p>Mistral reports a word error rate of about 4% for Transcribe V2 on the FLEURS benchmark and says it runs about three times faster than ElevenLabs' Scribe v2 at roughly one-fifth of the cost. These are the company's own figures and haven't been independently verified.</p> <p>See more AI products in our <a href="https://ainewsround.com/blog/higgsfield-takes-ai-video-from-1-minute-to-over-10" rel="noopener">Tools &amp; Applications coverage</a>.</p>