Mistral Voxtral

Voxtral is Mistral’s family of open speech models. It transcribes audio and answers questions about it, handling up to 40 minutes of sound in one request, and it ships with downloadable weights.

What is Mistral Voxtral?

Voxtral is a speech model line from Mistral AI, released in 2025. It comes in two sizes, a 3B and a larger 24B variant, both open-weights under a permissive license. Unlike a plain transcriber, it can take audio plus a text prompt and reason over both, so you can ask what a meeting decided rather than just get a transcript.

Product Features

• Two sizes: Voxtral Small (3B) and Voxtral Mini, plus a larger 24B model.
• Long audio: up to 40 minutes per request.
• Audio Q&A: answers written questions about what is said in the clip.
• Multilingual: strong English, Spanish, French and more.
• Open weights: download and self-host, or call it via Mistral’s API.

Product Highlights

Open weights are the argument. You can run Voxtral on your own GPUs, fine-tune it, and keep audio inside your network — something closed transcription APIs cannot offer. The 40-minute context also beats many hosted services that cap uploads at a few minutes.

Realistic catch: the biggest model is heavy and needs a serious GPU. Speaker diarization, the “who said what” labeling, is not its main strength, so for meeting apps you may still pair it with a dedicated diarization tool.

Use Cases

• Transcribing long recordings without a minute-based cap.
• Self-hosting transcription for privacy-sensitive audio.
• Building a “summarize this call” feature on open weights.
• Multilingual transcription for European teams.
• Fine-tuning on domain jargon like medical or legal terms.

Is Mistral Voxtral free?

The weights are free to download under an open license, though Mistral’s own terms restrict some commercial reselling. Using it through Mistral’s API is usage-priced, and the exact per-minute rate changes, so check the current pricing page.

Quick Summary

• Voxtral is Mistral’s open speech-to-text model family.
• Handles up to 40 minutes of audio per request.
• Open weights, self-hostable, multilingual.
• Weak on speaker diarization.