Voxtral
By Mistral AI
Voxtral is Mistral AI's family of open-weight audio-understanding language models, released in 2025, capable of speech transcription, audio question-answering, and voice-driven function calling built on top of Mistral's text model…
Definition
Voxtral is Mistral AI's family of open-weight audio-understanding language models, released in 2025, capable of speech transcription, audio question-answering, and voice-driven function calling built on top of Mistral's text model architecture.
Overview
Voxtral extends Mistral AI's open-weight model lineup into the audio modality, following the company's earlier expansions into vision (Pixtral) and code (Codestral/Devstral). Released in two sizes — Voxtral Small and Voxtral Mini — the models are designed to natively understand spoken audio rather than relying on a separate speech-to-text pipeline feeding transcribed text into a text-only LLM. This lets Voxtral handle tasks like transcription with strong accuracy across multiple languages, answering questions about the content of an audio clip, summarizing long audio recordings such as meetings or calls, and triggering function calls directly from spoken instructions in voice-driven applications. Mistral positioned Voxtral as a competitor to both dedicated automatic speech recognition (ASR) systems like OpenAI's Whisper and to proprietary multimodal audio capabilities embedded in closed models like GPT-4o's voice mode and Gemini's audio understanding. A key differentiator Mistral emphasized was long-audio handling — Voxtral models were built to process substantially longer audio inputs (reportedly up to tens of minutes) than many comparably sized open speech models, useful for transcribing and reasoning over full meetings or lectures rather than short clips. Voxtral was released under the Apache 2.0 open-weight license for its smaller variants, continuing Mistral's pattern of releasing broadly usable open models across modalities, and is available via Mistral's API as well as for local/self-hosted deployment. It fits into the growing category of open audio-language models — alongside efforts from other labs — aimed at reducing dependence on closed, API-only voice AI products for developers building voice assistants, transcription tools, and audio-analysis applications.
Key Concepts
- Native audio understanding rather than a separate ASR-then-LLM pipeline
- Released in Voxtral Small and Voxtral Mini sizes
- Handles transcription, audio Q&A, summarization, and voice-driven function calling
- Designed to process long audio inputs (tens of minutes) more effectively than many peers
- Released under the Apache 2.0 open-weight license
- Available via Mistral's API and for self-hosted local deployment