Mistral: Voxtral Mini 3B 2507

Voxtral Mini 1.0 3B 2507 is Mistral AI’s compact text-and-audio understanding model built on the Ministral 3B language-model family. It accepts speech recordings and text instructions and produces text, including transcripts, translations, answers about the audio and structured summaries. The original Voxtral release positions this 3B model for compact or local applications, alongside the separate 24B Voxtral Small model. It is also distinct from the hosted transcription-optimised Voxtral Mini Transcribe service.

The model has a dedicated transcription mode that automatically predicts the source language by default and transcribes accordingly. Choose that mode when the required artifact is the spoken wording. Audio understanding instead permits a direct question or summary request about a recording, without requiring a separate speech-recognition stage followed by a text model. An interview can therefore produce a full transcript, a translation or an answer to a narrow question; specify which result is needed so those different tasks do not become one ambiguous brief.

The exact release card lists a 32K-token context and audio up to 30 minutes for transcription or 40 minutes for understanding. These are different task-specific duration statements, not a shared audio allowance for every operation. Keep the relevant recording, instructions and conversation within the intended task scope. The maker’s named languages are English, Spanish, French, Portuguese, Hindi, German, Dutch and Italian, with automatic language detection. For a translation or answer, distinguish the recording’s language from the required output language.

Several audio inputs in a message and audio across multiple user turns are supported. Give each recording a clear identifier when asking for comparisons. An original example might compare two spoken explanations and return which requirements changed, with a separate column for each source. A follow-up can ask for the detail behind one finding or supply another recording. The card also demonstrates text-only and audio-only use, allowing the model to work with the written question as well as the speech evidence.

For a meeting summary, supply the recording and request a useful structure such as topics, stated decisions and follow-up actions. Keep uncertain names, dates or responsibilities separate from facts that are clearly spoken. For an interview, ask for the statement that answers a particular question and enough surrounding context to make it understandable. These are task designs for a usable text artifact; they do not turn general audio understanding into a guarantee of diarisation, word timestamps or identification of an unnamed speaker.

The manufacturer describes function calling from spoken intent. In such a workflow, the application supplies function definitions, the model proposes the relevant call and arguments, and the application executes it and returns the observation. A voice request can therefore guide a lookup or another tool-assisted task, but the model’s proposed call is distinct from an executed action. Preserve the actual returned result when continuing the conversation so a later answer reflects what happened rather than what was requested.

The released usage card recommends temperature 0.0 for transcription, and temperature 0.2 with top_p 0.95 for audio understanding and chat. It states that system prompts are not supported in this flow, so put the task and response requirements in the user input. Its examples decode recordings into audio inputs and include MP3 material, but they do not establish an exhaustive codec list or a separate hard maximum output length. The 500-token generation setting in the examples remains an example budget.

Prepare a clear recording containing the material relevant to the task and specify whether you need original wording, a translation, an answer or a summary. The card also demonstrates batched inference with several conversations, a producer usage pattern separate from the identity of an individual result. Preserve the 3B 2507 edition and its transcription-versus-understanding duration guidance; the later Mini Transcribe 2 model’s diarisation, 13-language profile and longer recording support are separate producer features.