Aayush9029/voxtral-mini-3b-8bit

Voxtral Mini 3B (MLX, 8-bit)

8-bit quantized MLX weights for Mistral's Voxtral Mini speech-to-text model, optimized for Apple Silicon inference. Recommended for most users — best balance of quality and download size.

Voxtral Mini is built on Ministral 3B with state-of-the-art audio understanding capabilities. It supports transcription, translation, Q&A, summarization, and function calling directly from audio input.

Features

• 8 languages: English, Spanish, French, Portuguese, Hindi, German, Dutch, Italian
• Long-form audio: up to 30 min transcription, 40 min understanding
• 32k token context
• Voice-triggered function calling

Specifications

Property	Value
Total Parameters	4.68B (8-bit quantized)
Precision	8-bit
Download Size	5 GB
License	Apache 2.0

Usage

Default model for Mac X — on-device speech transcription on Apple Silicon via MLX.

See also: Full precision (bfloat16) (17.4 GB) | 4-bit mixed quantized (3.2 GB)

License

Apache 2.0 — original model by Mistral AI.