Mistral AI Releases Voxtral: Open Source AI Audio Model

The Rise of Open-Source AI Audio Models

With the increasing sophistication of AI systems, voice interaction is rapidly becoming the primary method for communicating with machines. Responding to this trend, French AI startup Mistral has entered the audio technology arena with its inaugural open model.

Introducing Voxtral: A New Approach to Speech Intelligence

Mistral unveiled Voxtral on Tuesday, representing its first suite of audio models specifically designed for business applications. The company positions Voxtral as the pioneering open model capable of delivering genuinely practical speech intelligence for real-world deployment.

Historically, developers faced a trade-off. They could opt for inexpensive, open-source systems that often struggled with accurate transcriptions and comprehension, or choose high-performing, yet closed, solutions that came with increased costs and limited deployment control. Voxtral aims to eliminate this dilemma.

For organizations, this translates to a more affordable alternative, with Mistral asserting that Voxtral’s pricing is “less than half” that of comparable offerings.

Capabilities and Features of Voxtral

Mistral states that Voxtral can process audio recordings up to 30 minutes in length. Leveraging its LLM foundation, Mistral Small 3.1, the model demonstrates comprehension capabilities extending to 40 minutes of audio.

This allows users to engage with the audio content through question answering, summary generation, and the execution of voice-activated commands, such as API calls or function execution. Furthermore, Voxtral supports multiple languages, including English, Spanish, French, Portuguese, Hindi, German, Dutch, and Italian, for both transcription and understanding.

Voxtral Model Variants

The company is providing two distinct versions of its speech understanding models. Voxtral Small, equipped with 24 billion parameters, is intended for large-scale production deployments and is positioned as a competitor to ElevenLabs Scribe, GPT-4o-mini, and Gemini 2.5 Flash.

Voxtral Mini, featuring 3 billion parameters, is designed for local and edge computing environments. An exceptionally cost-effective, streamlined, and rapid API version of the 3 billion parameter model, named Voxtral Mini Transcribe, is also available. This version is optimized solely for transcription tasks and is projected to surpass OpenAI Whisper in performance at a lower cost.

Access and Pricing

Users can explore Voxtral’s capabilities without charge by downloading the API from Hugging Face or by experimenting with the models within Mistral’s chatbot, Le Chat. API integration into applications begins at a rate of $0.001 per minute, as indicated by the company.

Recent Developments at Mistral

This launch follows closely on the heels of Mistral’s announcement of Magistral, its first family of reasoning models, which employ a step-by-step problem-solving approach to enhance reliability.

Mistral, a leading AI company based in Europe, is widely recognized for its strong advocacy of open-source AI models. Recent reports from TechCrunch suggest the company is currently in discussions to secure up to $1 billion in equity funding from investors, including Abu Dhabi’s MGX fund.