Voxtral lets an application ask questions about recorded speech or summarize it without assembling separate transcription and language-model stages. Mistral AI's July 2025 release describes models that accept audio directly and produce transcriptions, answers, or summaries. It offers downloadable 3-billion- and 24-billion-parameter versions under Apache 2.0, with the smaller version positioned for local devices. That gives teams a route to running speech understanding on their own infrastructure, including applications where the desired output is an answer about a recording.
Terms in this piece · Glossary
open weights — A model whose trained parameters are published for anyone to download and run — unlike API-only models you can access but never possess.
