Voxtral-Mini-3B-2507

Multimodal audio understanding with transcription and reasoning

About model

Voxtral-Mini is an enhancement of Ministral 3B that incorporates state-of-the-art audio input capabilities while retaining best-in-class text performance. This compact 3B parameter model excels at speech transcription, translation, and audio understanding—delivering best-in-class ASR performance across the world's most widely spoken languages.

Audio Understanding

40min

32K context for long-form audio processing

Languages Supported

Automatic language detection & transcription

Compact Parameters

Efficient audio + text in small model size

Model key capabilities

Unified Audio-Text Model: No separate ASR pipeline needed—direct Q&A, summarization, and function calling on audio
Production-Ready Transcription: Dedicated mode with automatic language detection across 8 major languages
Long-Form Audio: Handle up to 40 minutes of audio for understanding tasks with 32K token context
Retains Text Excellence: Built on Ministral-3B backbone—maintains full text reasoning capabilities alongside audio

Quickstart guides

Audio

How to Build Real-time Audio Transcription App

Audio

Open NotebookLM: PDF to Podcast

Audio

Speech-to-Text Docs

API usage

cURL
Python
Typescript

Endpoint:

mistralai/Voxtral-Mini-3B-2507

curl -X POST "https://api.together.xyz/v1/audio/transcriptions" \
  -H "Authorization: Bearer $TOGETHER_API_KEY" \
  -F "model=mistralai/Voxtral-Mini-3B-2507" \
  -F "language=en" \
  -F "response_format=json" \
  -F "timestamp_granularities=segment"

from together import Together

client = Together()
response = client.audio.transcribe(
    model="mistralai/Voxtral-Mini-3B-2507",
    language="en",
    response_format="json",
    timestamp_granularities="segment"
)
print(response.text)

import Together from "together-ai";

const together = new Together();

const response = await together.audio.transcriptions.create(
  model: "mistralai/Voxtral-Mini-3B-2507",
  language: "en",
  response_format: "json",
  timestamp_granularities: "segment"
});
console.log(response)

Model card
Architecture Overview:
• Built on Ministral-3B backbone with added audio encoder capabilities
• 32K token context length enabling up to 30 minutes for transcription, 40 minutes for understanding
• Dual-mode operation: dedicated transcription mode and full audio understanding mode
• Native multimodal architecture combining audio and text processing
• Automatic language detection across 8 major world languages

Training Methodology:
• Enhanced from Ministral-3B language model with state-of-the-art audio capabilities
• Trained for leading performance on speech transcription benchmarks
• Optimized for multilingual understanding across English, Spanish, French, Portuguese, Hindi, German, Dutch, Italian
• Retains full text understanding capabilities from base Ministral-3B model
• Supports function-calling directly from voice input

Performance Characteristics:
• State-of-the-art Word Error Rate (WER) across FLEURS, Mozilla Common Voice, and Multilingual LibriSpeech benchmarks
• Competitive with specialized ASR models while adding audio understanding capabilities
• Outperforms models like Whisper large-v3, GPT-4o-mini-transcribe, and Scribe across multiple languages
• Dedicated transcription mode maximizes ASR performance
• Built-in Q&A and summarization without separate ASR preprocessing
• Compact 3B parameter size for efficient deployment
‍
Applications & use cases
Speech Transcription & ASR:
• State-of-the-art automatic speech recognition across 8 languages
• Automatic language detection—no manual language specification needed
• Long-form transcription up to 30 minutes of continuous audio
• Production-ready dedicated transcription mode for maximum accuracy
• Ideal for meeting transcription, podcast transcription, and content creation

Voice Assistants & Conversational AI:
• Direct Q&A from audio without separate ASR pipeline
• Function-calling straight from voice for backend integration
• Multi-turn audio conversations with context retention
• API calls triggered directly by spoken user intents
• Perfect for voice-enabled applications and smart assistants

Audio Understanding & Analysis:
• Built-in summarization of audio content up to 40 minutes
• Extract insights and answer questions about audio without transcription step
• Structured summary generation from voice recordings
• Meeting analysis, lecture comprehension, podcast summarization
• Content moderation and audio classification

Multilingual Applications:
• 8 language support: English, Spanish, French, Portuguese, Hindi, German, Dutch, Italian
• Automatic language detection eliminates manual configuration
• Cross-language transcription and translation capabilities
• Global customer service and support automation
• International content localization and accessibility

Enterprise & Production:
• Customer service call analysis and quality monitoring
• Voice-activated workflows and business process automation
• Accessibility solutions for hearing-impaired users
• Media and broadcasting transcription services
• Legal and medical audio documentation

Developer & Integration:
• Unified model eliminates complex ASR + LLM pipelines
• Direct audio-to-action workflows via function calling
• Retains full text capabilities for hybrid applications
• Easy deployment with vLLM and Transformers support
• Compact 3B size enables cost-effective scaling
‍

Related models

Model specifications

Model data

Model provider
Mistral AI
Type
Audio
Main use cases
Small & Fast
Speech-to-Text
Deployment
Serverless
Endpoint
mistralai/Voxtral-Mini-3B-2507
Parameters
3B
Context length
32K
Input modalities
Audio
Output modalities
Text

Released
July 1, 2025
Last updated
November 2, 2025
External link
Provider docs
Category
Transcribe

Run in Playground

Quickstart docs

Deploy model

Voxtral-Mini-3B-2507

About model

API usage

Model card

Applications & use cases