Mistral launches Voxtral, an open audio AI to challenge closed speech systems
French AI startup Mistral just dropped Voxtral, its first open speech intelligence model family for businesses. The company says Voxtral is the first open model “truly usable” in production, filling the gap between cheap, low-quality open speech AIs and expensive closed ones.
Voxtral costs less than half of comparable corporate speech solutions, aiming to give developers better transcription, understanding, and control without the corporate lock-in. 
Voxtral can transcribe up to 30 minutes of audio. Thanks to its Large Language Model backbone, Mistral Small 3.1, it can actually understand up to 40 minutes of audio, letting users ask questions, generate summaries, or turn voice commands into API calls and functions in real time. The model is multilingual, handling English, Spanish, French, Portuguese, Hindi, German, Dutch, and Italian.
Two versions are out now:
- Voxtral Small with 24 billion parameters for full production. Competes with ElevenLabs Scribe, GPT-4o-mini, and Gemini 2.5 Flash.
- Voxtral Mini with 3 billion parameters for edge and local use. Plus a stripped-down API version optimized for faster, cheaper transcription called Voxtral Mini Transcribe. This version claims to beat OpenAI Whisper at less than half the price.
Developers can try Voxtral for free by downloading the API on Hugging Face or testing Mistral’s chatbot Le Chat. API pricing starts at $0.001 per audio minute.
This launch follows Mistral’s July roll-out of Magistral, its reasoning models designed for more reliable step-by-step problem solving.
Mistral is one of Europe’s top AI startups and pushes open source AI models hard. TechCrunch recently revealed the company’s talks to raise $1 billion from investors like Abu Dhabi’s MGX fund.
Image Credits: Mistral
