Voxtral Mini Transcribe Realtime
Mistral AI · Francja · 2026
Open-weights live transcription in a 4B model — latency configurable below 200 ms, small enough to run on a local machine.
Voxtral Mini Transcribe Realtime, released alongside its batch sibling on 4 February 2026, transcribes audio as it arrives rather than after the recording ends. Latency is configurable down to under 200 milliseconds — the threshold below which a voice assistant stops feeling like it is waiting for you to finish and starts feeling like it is listening. The model is deliberately small: 4 billion parameters, which is what makes local and edge deployment realistic rather than aspirational. Mistral released the weights under an Apache-2.0 licence, so it can be run on a laptop or on hardware inside a building where the audio is not allowed to leave — a decisive argument in healthcare, legal and public-sector work. Through the API it is priced at $0.006 per minute of streaming audio, twice the batch rate. It covers the same 13 languages as the batch model, but not the same features: speaker diarization is not available in realtime mode, so a live transcript will not tell you who is speaking. Applications that need both typically stream for the live view and re-run the recording through Voxtral Mini Transcribe 2 afterwards.
▸Videos
No videos yet.
▸Reviews
No reviews yet. Be the first!