Voxtral Mini Transcribe Realtime
Mistral AI
Voxtral Mini Transcribe Realtime is Mistral's 4B-parameter streaming speech-to-text model, built from a roughly 3.4B language backbone and a roughly 970M causal audio encoder trained from scratch. It supports 13 languages and a configurable transcription delay, which the model card describes as multiples of 80 ms between 80 and 1200 ms plus a standalone 2400 ms setting; Mistral recommends 480 ms. Weights are released in BF16 under Apache-2.0, and the hosted version is served through the /v1/audio/transcriptions endpoint.
Strengths
- Natively streaming architecture with a custom causal audio encoder, giving a delay of under 500 ms
- Configurable transcription delay: any multiple of 80 ms between 80 and 1200 ms, plus 2400 ms as a standalone value
- At the recommended 480 ms delay it matches leading offline open-source transcription models on the vendor's own tables
- 13 languages: English, French, Spanish, German, Russian, Chinese, Japanese, Italian, Portuguese, Dutch, Arabic, Hindi and Korean
- Sliding window attention in both the audio encoder and the LLM backbone, which the card says allows 'infinite' streaming
- 4B parameters (about 3.4B language model plus about 970M audio encoder), released in BF16 under Apache-2.0 for research and commercial use
- Hosted by Mistral at /v1/audio/transcriptions for $0.006 per minute
Caveats
- Accuracy is below Mistral's own offline Voxtral Mini Transcribe 2.0 on every FLEURS language: 8.72% average WER at 480 ms versus 5.90% offline.
- Arabic is by far the weakest language, at 22.53% WER at 480 ms against 13.54% for the offline model; Korean is 15.74%.
- Transcription only. The vendor describes it as 'optimized for live transcription purposes only', so it is not an audio-understanding or conversational model.
- Low delay settings degrade sharply: average FLEURS WER rises to 12.60% at 160 ms and 10.80% at 240 ms.
- No context window or maximum output length is published (the docs page shows '--'); the 131072 figure in the card is a recommended vLLM --max-model-len default, not an advertised context window. The card is also internally inconsistent about the delay floor, saying 240 ms in one place and 80 ms in another.
- Vendor
- Mistral AI
- License
- Apache-2.0
- Release Date
- 2026-02-04
- Modalities
- audiotext
| Benchmark | Score | Max |
|---|---|---|
| FLEURS (average word error rate, 480 ms delay) | 8.7 | 100.0 |
| TEDLIUM (long-form English word error rate, 480 ms delay) | 3.2 | 100.0 |
Capabilities
Resources
Pricing
Mistral's model page prices the hosted model at $0.006 per minute of audio (EUR 0.0053 per minute), with no separate output charge. The BF16 weights are published under Apache-2.0 at mistralai/Voxtral-Mini-4B-Realtime-2602 for self-hosting.