Voxtral is Mistral AI's family of speech models: they turn spoken audio into text, answer questions about it, and in newer versions transcribe live and speak.
In plain words
Voxtral is Mistral AI's line of speech models. The first two came out on 15 July 2025: a 24-billion-parameter model for large services and a 3-billion one for local or small devices, both under the Apache 2.0 licence. Besides writing down what was said, they answer questions about a recording, summarise it and detect its language. Voxtral Mini 4B Realtime transcribes 13 languages while the speaker is still talking, and an Arabic-dialect version followed. Voxtral 3 models also label who is speaking and talk in preset or cloned voices.
A quick example
Imagine a school that streams its assemblies online. A realtime speech model listens to the microphone feed and puts captions on screen a moment behind the speaker, so students who cannot hear well can follow along.
Where you'll meet it
- Live captions and voice assistants.
- Transcribing recordings and labelling who spoke.
Related
Sources
- DocsVoxtral, Mistral AI · read 12 Oct 2026
- Docsmistralai/Voxtral-Mini-4B-Realtime-2602, Mistral AI (Hugging Face) · read 9 Oct 2026
- Docsmistralai/Voxtral-Mini-4B-Realtime-Arabic, Mistral AI (Hugging Face) · read 9 Oct 2026
- DocsAudio, Mistral Docs · read 20 Oct 2026
Checked by WhatIsAI against the sources above · 20 Oct 2026. A full five-minute explainer with a diagram may follow. Spotted an error? Report it.