Mistral: Voxtral Mini Transcribe

Mistral Voxtral Mini Transcribe is a speech-to-text model from Mistral, derived from the Voxtral Mini family. It is built specifically for transcription work rather than general chat or generation, and it sits in the transcription category on api14. If you already know Mistral for its language models, this is the same provider applying that work to audio, with a narrower and more predictable job: turning recorded speech into written text.

The model takes audio as its only input and returns a transcription as its output. It does not accept text prompts or images, so there is no multimodal mixing to plan around. The context window is 16,384 tokens, and the maximum output is 13,107 tokens. That output ceiling matters in practice because it caps how much transcribed text a single request can return, which in turn bounds how much audio you should send in one call. Long recordings may need to be split into smaller segments so each response stays within that limit.

Because the input is audio and the output is text, the obvious uses are meeting transcription, interview notes, voice memo cleanup, lecture capture, and podcast or video subtitling workflows where you need a written transcript before any further processing. It also fits pipelines that feed transcripts into downstream steps such as summarisation, search indexing, or keyword extraction, since the output is plain transcribed text. Anywhere you have recorded speech and need it as text, this model is the direct fit.

You call it through api14's OpenAI-compatible API using the model id "mistralai/voxtral-mini-transcribe". Because the interface follows the OpenAI transcription format, you send audio in the request and receive the transcript in the response, which means existing OpenAI transcription clients usually need only the base URL and model id changed. The model lists no supported parameters, so there are no sampling or generation controls to tune; keep requests simple and rely on the defaults. If you need behaviour beyond plain transcription, this model is not the right tool, since it produces transcription output only and exposes no adjustable parameters.