Voxtral
modelYour notes
Mistral's speech model family. Voxtral Mini (3B) and Small (24B) handle long-form audio understanding (30-40 minutes). Voxtral TTS (4B) adds zero-shot voice cloning. Natively multilingual. Apache 2.0.
Voxtral 3 (20 October 2026) is the third generation and repositions the line around voice agents. It ships four models and a voice agent SDK that runs listening, reasoning, tool calling and speaking in one real-time streaming pipeline. The models are Voxtral Small Transcribe 3, Voxtral Mini Transcribe 3 ("4B parameters", batch transcription at $0.0035/min), Voxtral Mini Realtime 3 (sub-500 ms streaming, 35+ languages) and Voxtral Mini TTS 3 (35+ languages, 50+ dialects, $13 per 1M characters). Transcription now covers 50 languages, including all 24 official EU languages, with diarization, context biasing and native end-of-turn detection. Only the Mini sizes are disclosed. Mistral offers the family through its EU-hosted API or for self-deployment, but no Voxtral 3 weights were on Hugging Face at filing, so treat it as commercially licensed rather than open-weight.
The comparisons are Mistral's own. On diarization (cpWER averaged over AMI, Switchboard, CallHome and AISHELL-4) Small Transcribe 3 reports 18.4% and Mini Transcribe 3 18.6%, against Pyannote 19.0%, Microsoft AI Transcribe 2 20.3%, Scribe v2 23.0%, the previous Voxtral Mini Transcribe 2 30.9% and Gemini 3.5 Transcribe 38.4%. That makes diarization the clearest step over the last generation. On a 45-language FLEURS macro-average Small Transcribe 3 is around 4% WER, roughly tied with Microsoft AI Transcribe 2. Mistral also plots Artificial Analysis's speech-to-text WER index, where both Transcribe 3 models sit in the leading group but not first.
Voxtral-Mini-4B-Realtime-Arabic (Hugging Face, 8 October 2026, Apache 2.0) is an open streaming ASR model fine-tuned from Voxtral-Mini-4B-Realtime-2602 (about 4.4B parameters) for Modern Standard Arabic and 15 regional varieties, including code-switched speech. It was co-developed with Morocco's Ministry of Digital Transition (MTNRA) under the January 2026 Mistral–Morocco partnership. At a 480 ms delay it reports an average 8.82% CER across seven Arabic benchmarks, within 0.91 points of the offline Voxtral Transcribe Arabic's 7.91%.
Model Details
Variants
| Name | Parameters | Notes |
|---|---|---|
| Voxtral Mini 3B (2507) | — | audio understanding, open weights, Apache 2.0 |
| Voxtral Small 24B (2507) | — | audio understanding, open weights, Apache 2.0 |
| Voxtral 4B TTS (2603) | — | zero-shot voice cloning, open weights |
| Voxtral Mini 4B Realtime (2602) | — | streaming speech-to-text, open weights |
| Voxtral-Mini-4B-Realtime-Arabic | — | 2026-10-08; fine-tune of Voxtral-Mini-4B-Realtime-2602 (~4.4B) for MSA + 15 regional varieties; 8.82% avg CER at 480 ms on 7 Arabic benchmarks; Apache 2.0; co-developed with Morocco's MTNRA |
| Voxtral Small Transcribe 3 | — | 2026-10-20; size undisclosed; 50 languages; diarization cpWER 18.4%; API / self-deployment, no public weights |
| Voxtral Mini Transcribe 3 | 4B | 2026-10-20; batch transcription, $0.0035/min; cpWER 18.6%; no public weights |
| Voxtral Mini Realtime 3 | — | 2026-10-20; sub-500 ms streaming transcription, 35+ languages; no public weights |
| Voxtral Mini TTS 3 | — | 2026-10-20; 35+ languages, 50+ dialects, $13 per 1M characters; no public weights |