Mistral Voxtral Mini Transcribe Realtime
Voxtral Mini Transcribe Realtime is Mistral AI's streaming speech-to-text model. Mistral's model documentation gives the API id voxtral-mini-transcribe-realtime-2602, a release date of February 4, 2026, and describes it as an efficient audio-input model pre-trained and optimized for live transcription only. It launched with the batch model Voxtral Mini Transcribe V2 as part of Voxtral Transcribe 2. The model is current as of 2026-10-29, with no deprecation shown on the model page.
Model Specs
| Spec | Voxtral Mini Transcribe Realtime |
|---|---|
| API id | voxtral-mini-transcribe-realtime-2602 |
| Release | February 4, 2026 |
| Open weights | Voxtral-Mini-4B-Realtime-2602 on Hugging Face |
| Size | About 4B parameters: roughly 3.4B language model and 970M causal audio encoder, BF16 |
| Languages | 13: English, Chinese, Hindi, Spanish, Arabic, French, Portuguese, Russian, German, Japanese, Korean, Italian, Dutch |
| License | Apache 2.0 |
| API price | $0.006 per minute |
| Hardware | A single GPU with at least 16 GB of memory, per the Hugging Face card |
Key Features
- Configurable latency: the Hugging Face card says the transcription delay can be set from 240 ms to 2.4 s, and recommends 480 ms as the best balance of accuracy and latency. Mistral's announcement says delay is configurable down to sub-200 ms, and gives accuracy figures only at 480 ms and 2.4 s.
- Accuracy trade-off, per Mistral: at 480 ms the average FLEURS word error rate is 8.72%, compared with 5.90% for the offline Voxtral Mini Transcribe 2.0. On long-form English it is 5.05% versus 4.08%. These are Mistral's own numbers.
- Open and edge-friendly: the Apache 2.0 license allows research and commercial use, and Mistral describes it as deployable on edge devices.
- Streaming runtime: vLLM is the recommended runtime through its realtime streaming endpoint, and the card also lists Transformers 5.2.0 or later.
Use Cases
- Live captions and subtitles for meetings, streams and events.
- Voice assistants that need transcription as the user speaks.
- Self-hosted transcription where audio cannot leave your infrastructure.
A limitation: it is for live transcription only. For batch jobs up to three hours long, Mistral points to Voxtral Mini Transcribe V2, which is priced at $0.003 per minute. See Voxtral Mini Transcribe 2.
Getting Started
- For the API, call
voxtral-mini-transcribe-realtime-2602through Mistral's audio endpoints, as described in the model docs. - To self-host, download the weights from the Hugging Face card and serve them with vLLM, using temperature 0.0.
- Start with a 480 ms delay and adjust for your latency needs.
- Test on your own audio and languages before you ship.
FAQ
What is Voxtral Mini Transcribe Realtime?
Mistral's open-weight, streaming speech-to-text model, released February 4, 2026.
How much does it cost?
$0.006 per minute through Mistral's API, according to the model docs. Self-hosting the Apache 2.0 weights has no per-minute fee.
Which languages does it support?
Thirteen, including English, Chinese, Hindi, Spanish, Arabic, French, Portuguese, Russian, German, Japanese, Korean, Italian and Dutch.
How low is the latency?
The Hugging Face card gives a configurable delay of 240 ms to 2.4 s, while Mistral's announcement says it can go down to sub-200 ms.
Alternatives
- Voxtral Mini Transcribe 2: Mistral's batch transcription model.
- Voxtral TTS: Mistral's text-to-speech model.
- Whisper large-v3: OpenAI's open speech recognition model.
- Deepgram Nova-2: a hosted speech-to-text model.
Conclusion
Voxtral Mini Transcribe Realtime is a practical choice for live transcription when you want open weights and a clear latency dial. Read Mistral's model docs for the current API details, and see more in the models category.
Sources
- Websitehttps://docs.mistral.ai/models
- Social Mediahttps://x.com/MistralAI
Comments
No comments yet. Be the first to comment!