Voxtral

What it is

Voxtral is a family of speech-understanding models released by Mistral AI on 15 July 2025 in two sizes: Voxtral Small, with 24 billion parameters, and Voxtral Mini, with 3 billion. Both were released under the Apache 2.0 licence.^[1] Audio goes in and text comes out, but the family was not built as a transcription tool alone: the same model is meant to answer questions about a recording, summarise it, or act on what the speaker asks for.^[1]

Where it came from

Voxtral is built on Mistral Small 3.1 and keeps that model's text abilities.^[1] On the day of release Mistral described three variants: Voxtral Small and Voxtral Mini for chat use cases, and Voxtral Mini Transcribe, which was optimised for transcription.^[2]

Long audio in one context

The models take a 32K-token context. That covers up to 30 minutes of audio when the task is transcription, and up to 40 minutes when the task is understanding the audio.^[1] Long inputs matter here because a single recording often needs to be held in view at once — a meeting, a lecture, a call — rather than cut into fragments that the model sees one at a time.

Beyond transcription

Voxtral can answer questions about audio content or generate structured summaries of it.^[1] It can also trigger backend functions or API calls from spoken user intents.^[1] That last capability points at voice-driven software: instead of a person tapping through menus, they say what they want, and the model decides which function should run.

Reported results

On transcription, Mistral said Voxtral outperforms Whisper large-v3, GPT-4o mini Transcribe and Gemini 2.5 Flash.^[1] The company also said the models beat GPT-4o mini Transcribe and Gemini 2.5 Flash across all tasks.^[1] These are vendor-stated comparisons.

Price and first deployment

Mistral offered Voxtral through an API with pricing starting at $0.001 per minute of audio.^[1] Two days after launch, from 17 July 2025, Voxtral powered the voice mode of Le Chat.^[4]

Later versions

On 4 February 2026 Mistral added Voxtral Mini Transcribe 2 and Voxtral Realtime, and on 20 October 2026 it introduced the Voxtral 3 family, covering transcription and text-to-speech.^[2] Voxtral Mini, listed under the model name voxtral-mini-2507, was given a deprecation date of 27 February 2026 and a retirement date of 31 May 2026, with Voxtral Mini Transcribe 2 named as its alternative.^[3]

Where it sits

The line of releases shows a family that started as three variants on one day — two chat-oriented sizes and one transcription-focused model — and then split further, into a transcription upgrade and a realtime model, before expanding into a family that also produces speech.^[2] For people building with it, the appeal was the combination: audio transcription and audio understanding, question answering, summaries and voice-triggered function calls in one set of models, under an open licence and priced by the minute.^[1]