Voxtral Realtime has Apache 2.0 weights; the batch model adds features for turning recordings into usable transcripts.
What happened
Voxtral Mini Transcribe V2 includes speaker diarization, word timestamps and context biasing across 13 languages. Mistral also introduced an audio playground and described configurable low latency for the real-time model.
Why it matters
A meeting transcript and a live voice assistant have different requirements. Speaker attribution improves navigation through recordings, while low delay affects whether spoken interaction feels responsive.
The bigger picture
Our take. Transcription should be tested with representative microphones, accents and background noise. Mistral's speed and accuracy claims are developer-reported, and the most useful metric may be whether important names and instructions survive intact.
Source checked November 18, 2026. Company claims are attributed; paragraphs marked “Our take” are editorial interpretation.