Voxtral Small (latest)

Voxtral Small is the audio-understanding member of Mistral AI's Voxtral family, designed for conversational interaction and question-answering over speech rather than straightforward transcription. It exposes this audio capability through a chat completions-style endpoint, letting developers mix spoken and written input in the same request and receive plain text replies. Because the model is released as open weights under the permissive Apache-2.0 license, teams can self-host it, fine-tune it on their own audio corpora, or build private voice assistants without being locked into a proprietary stack, while still benefiting from Mistral's research lineage in efficient dense transformers.

At its core the model is a 24-billion-parameter dense transformer with a 32,000-token context window, giving it enough room to follow long spoken instructions or to combine extended audio transcripts with surrounding text passages. Its text-token pricing is published, while audio input is metered separately, so costs scale sensibly with the amount of speech actually processed. In practice this combination of an open-weight license, a reasonably sized dense architecture, and genuine audio understanding makes Voxtral Small a practical fit for organizations that want a speech-aware assistant they can run on their own infrastructure for internal tooling, customer support, or domain-specific voice agents.