Mistral AI releases Voxtral speech models

Background

A speech model takes a recording and turns it into text. Two jobs are often treated separately: transcription, which writes down what was said, and understanding, which means answering a question about a recording or condensing it into a summary. A system that does only the first leaves the second to a separate language model, with the transcript passed along as an intermediate step.

What happened

On 15 July 2025 Mistral AI released two speech models, Voxtral Small and Voxtral Mini ^[1]. The two differ in size: Small has 24 billion parameters and Mini has 3 billion ^[1]. Parameters are the learned numbers inside a model; more of them generally mean a larger model that costs more to run, fewer mean something lighter and cheaper to deploy. Both models are licensed under Apache 2.0 ^[1], a permissive open-source licence whose terms allow use, modification and redistribution.

Mistral also made the models available through an API, with pricing starting at $0.001 per minute of audio ^[1].

Key details

For transcription, both models handle up to 30 minutes of audio ^[1]. That is enough for a lecture, a meeting or a long interview to be processed in one pass rather than in short chunks stitched together afterwards.

They are not limited to producing a transcript. According to Mistral, they can answer questions about the content of a recording and generate structured summaries of it ^[1] — turning a conversation into organised notes, for instance, rather than a flat wall of text. Doing this inside the speech model means the audio does not have to be transcribed first and then handed to a separate system for the follow-up question.

Mistral said Voxtral comprehensively outperforms Whisper large-v3, the model it named as the comparison point ^[1].

Why it matters

The release put a speech model under a permissive licence in two sizes at once ^[1]. For anyone building on top of it, choosing between the 24-billion-parameter model and the 3-billion-parameter one is a choice about how much computation to spend per hour of audio, and the Apache 2.0 terms leave room to run the weights on their own machines as well as to call them remotely.

The published API rate of $0.001 per minute gives the same models a per-minute price ^[1], so they can be used as a hosted service or as something to download, run and modify directly. Speech transcription is usually billed and planned for by the minute of audio, which makes a per-minute figure the number a team compares when deciding what to use.

The comparison with Whisper large-v3 places the release against a named transcription baseline ^[1]. A benchmark claim of that kind is a starting point rather than an answer: whether the advantage holds across languages, accents and noisy recordings is what users tend to check on their own material.