Mistral: Voxtral Small 24B 2507 STT

Mistral: Voxtral Small 24B 2507 STT is a speech transcription model from Mistral AI. It sits in the transcription category, and its name reflects both the larger 24B capacity of the model and its focus on speech to text work. Mistral positions it for transcription, translation, and audio understanding workloads that benefit from that larger model capacity, so it is aimed at developers who need more than a minimal speech to text pass over their audio.

The model takes audio as its input modality and returns a transcription as its output. That is the whole shape of the interface: you send audio, you get text back. No context length is listed for this model, and no supported request parameters are listed either, so you should treat it as a focused single purpose endpoint rather than a general chat model you can steer with sampling controls. The details also list a max output of 0, which means there is no separately advertised output token budget to plan around.

The practical uses follow directly from that audio in, transcription out shape. Meeting and call recordings can be turned into text for search, review, or archival. Voice notes and interviews can be transcribed in bulk. Because the provider description also names translation, the model can be used where the goal is not only to capture what was said but to render it in another language. Audio understanding workloads, such as extracting the content of spoken material for downstream processing, are another fit. If your pipeline needs a text model to summarise or classify the transcript afterwards, that is a separate step with a separate model.

You call it through api14's OpenAI compatible API using the model id "mistralai/voxtral-small-24b-2507-stt". Because the endpoint follows the OpenAI format, you can point existing OpenAI client libraries at api14 and send your audio in the request, then read the transcription from the response. No supported parameters are listed for this model, so there is nothing to tune beyond selecting the model and supplying the audio; keep your request simple and avoid assuming that temperature, top_p, or similar knobs will have an effect.

Since no context length is published and no parameters are exposed, this model is best treated as a dedicated transcription component rather than something you configure heavily. If you need a model that accepts text prompts alongside audio, or one where you can control generation behaviour through request parameters, another model in the catalogue may fit better. For straightforward speech to text and speech translation work, this one is the direct choice.