As of: 2026-10-25 23:49 CET.
Researched and written by AI agents (Claude Opus 5.5 in Claude Code), with a human editor. How we work: https://postcutoff.com/about/
Canonical page: https://postcutoff.com/m/voxtral-mini-transcribe-3/

# Voxtral Mini Transcribe 3 (+ Mini Transcribe Realtime 3)

Mistral AI, Voxtral family. Status: preview. Type: speech and audio. Released: 2026-10-20.

## How to call it

| Provider | Model id | Endpoint | Docs |
|---|---|---|---|
| Mistral API (batch) | `voxtral-mini-transcribe-3-0` | https://api.mistral.ai/v1/audio/transcriptions | https://docs.mistral.ai/models/voxtral-mini-transcribe-3-0 |
| Mistral API (realtime, WebSocket) | `voxtral-mini-transcribe-realtime-3-0` | https://api.mistral.ai/v1/audio/transcriptions | https://docs.mistral.ai/models/voxtral-mini-transcribe-realtime-3-0 |

## Pricing

- Per minute, batch: $0.0035
- Per minute, realtime: $0.008
- Unit as published: USD per minute of audio (voxtral-mini-transcribe-3-0 batch / voxtral-mini-transcribe-realtime-3-0 streaming)
- Source: https://docs.mistral.ai/models/voxtral-mini-transcribe-3-0
- Last checked: 2026-10-25

## Specs

- Input: audio
- Output: text
- Context window: 32,000 tokens
- Max output: not published
- Open weights: no
- Licence: proprietary

## Where it stands

- No newer speech and audio model in the registry yet.
- Before it from Mistral AI: Voxtral TTS (2026-03-23) https://postcutoff.com/m/voxtral-tts/
- Closed weights; access is hosted only.

## What stands out

- Compact 4B transcriber with diarization: 4B-parameter model for 50+ languages with speaker diarization (18.6% macro-average cpWER on Mistral's chart) and timestamps; 'small enough to be deployed in resource-constrained environments'. Source: https://mistral.ai/news/voxtral-3
- #5 on Artificial Analysis AA-WER (2.2%) (found after launch): Fifth on the Artificial Analysis non-streaming speech-to-text leaderboard (AA-WER 2.2%, $3.50 per 1,000 min), tied in WER with ElevenLabs Scribe v2, read 2026-10-25. Source: https://artificialanalysis.ai/speech-to-text
- Realtime variant with configurable latency and end-of-turn detection: Voxtral Mini Transcribe Realtime 3 streams over one WebSocket session with latency configurable from 80 ms to 2,400 ms, built-in end-of-turn detection (acoustic + semantic), context biasing and speaker diarization. Source: https://docs.mistral.ai/models/voxtral-mini-transcribe-realtime-3-0

## Notes

Artificial Analysis speech-to-text (https://artificialanalysis.ai/speech-to-text), read 2026-10-25: Voxtral Mini Transcribe 3 rank #5, AA-WER 2.2%, $3.50 per 1,000 minutes (Realtime 3 not in that non-streaming ranking). Docs: v3.0, 'Public Preview', 'Premier' tag; no open weights found (unlike the Apache 2.0 Voxtral Realtime 2602 of Feb 2026). Realtime language count is inconsistent on Mistral's page: 35+ in the post body, 40+ on its card. Realtime 3 is also called 'Voxtral Realtime Transcribe 3'. Sibling: voxtral-small-transcribe-3 ($0.004/min, higher accuracy). Succeeds voxtral-transcribe-2 (voxtral-mini-2602 / voxtral-mini-transcribe-realtime-2602).

Cheaper Voxtral 3 transcription model, plus its streaming sibling for voice agents.

```bash
curl https://api.mistral.ai/v1/audio/transcriptions -H "Authorization: Bearer $MISTRAL_API_KEY" \
  -F model=voxtral-mini-transcribe-3-0 -F file=@audio.mp3
```

Sources: https://mistral.ai/news/voxtral-3 · https://docs.mistral.ai/models/voxtral-mini-transcribe-3-0 · https://docs.mistral.ai/models/voxtral-mini-transcribe-realtime-3-0 · https://artificialanalysis.ai/speech-to-text

Model page: https://postcutoff.com/m/voxtral-mini-transcribe-3/