Mistral: Voxtral Small 24B 2507

Mistral: Voxtral Small 24B 2507 is a model from Mistral, released by the provider on 2025-10-30. It sits in the text category and is built as an enhancement of Mistral Small 3, adding audio input capabilities while retaining the text performance of that earlier model. If you already know Mistral Small 3, the useful way to think about Voxtral Small is that it keeps the text behaviour and extends what the model can accept as input.

On inputs and outputs, the model accepts text, audio and files, and it produces text only. The context window is 32768 tokens, and the maximum output is 26214 tokens. According to the provider, it excels at speech transcription, translation and audio understanding, which means audio is not a bolted-on extra but a first-class input path alongside text. The combination of a 32k context and a large output ceiling makes it practical for tasks where the input is long, such as a meeting recording plus instructions, and the answer needs room to breathe.

Those capabilities point to a fairly specific set of use cases. Speech transcription is the obvious one, whether you are turning recordings into written records or feeding spoken content into a text pipeline. Translation follows from the same audio understanding, for example converting speech in one language into text in another. Audio understanding covers the broader cases: summarising a call, extracting action items from a recording, or answering questions about what was said. Because text input is still supported and the model retains strong text performance, you can also mix modalities in one request, passing a transcript or a document alongside audio and asking for a combined result.

You call the model through api14's OpenAI-compatible API using the model id "mistralai/voxtral-small-24b-2507". The supported request parameters are frequency_penalty, max_tokens, presence_penalty, response_format, seed, stop, structured_outputs, temperature, tool_choice, tools and top_p. Tool calling and structured outputs are both available, so you can wire the model into an agent loop or force a JSON-shaped response when the output feeds another system. Sampling controls (temperature, top_p, the two penalties, seed and stop) behave as you would expect from an OpenAI-compatible endpoint.

The main constraint to keep in mind is the output modality: the model returns text, not audio, so it is not a speech synthesis option. The context window of 32768 tokens is also moderate, and very long documents or multi-hour recordings may need chunking before they fit. If your task is purely text and does not benefit from audio input, a text-only model may be a simpler fit. If you need audio in and text out, this model is aimed squarely at that.