Voxtral Mini 3B

Voxtral Mini 3B (mistralai/Voxtral-Mini-3B-2507) is the compact entry in Mistral's Voxtral speech family, released July 2025. It combines a Whisper large-v3 audio encoder with a Ministral 3B decoder, giving it transcription, translation, and genuine audio understanding - it can answer questions about audio, not just transcribe it - in a ~9.5GB footprint.

It supports 32K context (up to 30 minutes of audio for transcription, 40 for understanding), 8+ languages with automatic detection, multi-audio messages, and even function-calling triggered by spoken intent. The weights are Apache 2.0 and it runs in vLLM, SGLang, and llama.cpp with official GGUF builds.

Honest limits: it is speech-to-text, not speech generation - the TTS side of the Voxtral brand (Voxtral 4B TTS) is a separate model under CC BY-NC 4.0. Like all generative ASR models it can 'clean up' audio, which inflates word-error-rate on verbatim benchmarks.