Voxtral Small 2507
Voxtral Small is Mistral's 24B audio understanding model from July 2025. It remains in the current catalog with API access and Apache 2.0 weights. It can answer questions about speech, summarize recordings and produce text from audio input.
Ask about the recording directly
The launch describes up to 30 minutes of audio for transcription, or 40 minutes for understanding. It also describes function calling from spoken instructions. Voxtral retains the text abilities of its language-model base, so it can combine an audio recording with a text question.Voxtral release and audio understanding
Use it when the source audio matters
Practical guidance. A direct audio question can avoid losing details in an intermediate transcript. It still needs evidence checks. For an action request, verify the people, dates and conditions against the recording before the application executes a tool. Local deployment also requires an audio-capable runtime, not only the language weights.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Launch date | |
| Access | API, Downloadable |
| Model identifiers | voxtral-small-2507Voxtral-Small-24B-2507 |
| Context window | 32,000 tokens The model page states 32k context for audio and text interactions. Voxtral Small 2507 model page |
| Initial release stage | General availability.Voxtral Small 2507 model page |
| Language-model ancestry | Mistral Small 3.1Voxtral release and audio understanding |
| License | Apache 2.0Voxtral Small 2507 model page |
Related models
- Voxtral Small Transcribe 3
Mistral AI · Voxtral
A preview speech model that combines transcription and diarization with speech translation and audio question answering.
Launch date PreviewAPI - Voxtral Mini Transcribe 3
Mistral AI · Voxtral
A preview speech-to-text model for recorded audio with timestamps and speaker diarization.
Launch date PreviewAPI