Voxtral Small

Voxtral is Mistral AI's first open-weight audio model family, released July 15, 2025. It comes in two sizes: Voxtral Small at 24B parameters for production deployments, built on the Mistral Small 3.1 backbone, and Voxtral Mini at 3B for edge and local use. Both are released under Apache 2.0.

Voxtral is a speech-understanding model, not a speech generator: spoken audio and text go in, text comes out. It transcribes, translates, answers questions about recordings, summarizes conversations, and triggers function calls from voice instructions, with a 32K context window covering about 40 minutes of audio. Mistral reports state-of-the-art transcription and translation among open and closed models in its price class.

The honest caveats: it comprehends speech but does not synthesize it, so it is not a voice agent voice, and the benchmark claims are vendor-reported. A transcription-only API variant, Voxtral Mini Transcribe, is also offered.