Voxtral Realtime

Mistral's Apache 2.0 4B streaming speech-to-text model: 13 languages and a delay you set between 80 ms and 2.4 s.

In short. Voxtral Realtime writes down speech as it is spoken, instead of waiting for a recording to finish. It supports 13 languages, and you choose how much delay to accept: a shorter delay reacts faster, a longer one is more accurate. Mistral recommends 480 milliseconds as a balance, which suits voice assistants and live subtitles.

Summary

On 4 February 2026 Mistral AI launched Voxtral Transcribe 2, two speech-to-text models. Voxtral Mini Transcribe V2 is for batch transcription, and Voxtral Realtime is for live use. Only Voxtral Realtime has open weights, released under Apache 2.0 as Voxtral-Mini-4B-Realtime-2602. It has about 4.4B parameters: a roughly 3.4B language-model decoder and a roughly 970M audio encoder that was trained from scratch with causal attention, so it can stream. Both parts use sliding-window attention, which allows effectively unlimited streaming. It transcribes 13 languages: English, Chinese, Hindi, Spanish, Arabic, French, Portuguese, Russian, German, Japanese, Korean, Italian and Dutch.

The delay is a setting. You can pick multiples of 80 ms from 80 ms to 1,200 ms, or 2,400 ms, and the card recommends 480 ms as the sweet spot between latency and accuracy. In the card's own FLEURS table the average word error rate is 8.72% at 480 ms and 6.73% at 2,400 ms, against 5.90% for the offline Voxtral Mini Transcribe 2.0, so a longer delay narrows the gap. These are Mistral's numbers. The BF16 weights are about 8.9 GB, the card says the model is built for on-device deployment, and one text token equals 80 ms of audio.

The card recommends serving the model with vLLM over websocket streaming sessions, with the temperature at 0. The default maximum length of 131,072 tokens covers about three hours of continuous audio. Hugging Face showed about 1.5 million downloads in the 30 days to 11 November 2026. Mistral's blog lists speaker labels, word-level timestamps and context biasing as features of Voxtral Mini Transcribe V2, which is API-only, and the open model's card does not list them.

Good for

  • Live captions, voice assistants and meeting transcription that must appear while someone is speaking
  • Private or on-device deployments, since the model has 4B parameters and an Apache 2.0 license
  • Trading latency against accuracy to suit each use case
  • Speech in 13 widely spoken languages

Watch out for

  • Shorter delays raise the error rate, and the card's own table shows 12.60% on average at 160 ms against 8.72% at 480 ms
  • Error rates differ a lot by language, with the card showing 22.53% for Arabic, 15.74% for Korean and 4.90% for English at 480 ms
  • Speaker labels, word timestamps and context biasing are listed for the API-only Voxtral Mini Transcribe V2, not for this model
  • The card recommends vLLM with websocket streaming, so it needs more setup than a simple file-transcription tool

Performance claims come from the developer and have not been independently verified.

How much memory does it need?

Memory figures for each variant of Voxtral Realtime. Image, video, speech and music pipelines also load a text encoder and a decoder, so the memory needed can be larger than the main checkpoint.

Variants Parameters Download size Memory needed Fits on
Voxtral Mini 4B Realtime BF16 4.43B 8.86 GB 10 GB Laptop or phone

Download size covers every component needed to run the model, such as the generator, text encoder and decoder. Memory needed is a stated or published measured minimum where one exists; otherwise it is the download size plus about 15%. Larger images, longer clips and higher resolutions need more.

The license in plain words

One of the most widely used permissive licenses. You may use, change and sell products built on the model with only light conditions.

More about Apache 2.0

See also

Frequently asked questions

What is Voxtral Realtime?
Voxtral Realtime writes down speech as it is spoken, instead of waiting for a recording to finish. It supports 13 languages, and you choose how much delay to accept: a shorter delay reacts faster, a longer one is more accurate. Mistral recommends 480 milliseconds as a balance, which suits voice assistants and live subtitles.
Can I use Voxtral Realtime commercially?
Yes. Voxtral Realtime is released under Apache 2.0, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
Is Voxtral Realtime open source?
Voxtral Realtime is released under Apache 2.0, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
Can I run Voxtral Realtime on my own computer?
Yes. The smallest variant, Voxtral Mini 4B Realtime, needs about 10 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
When was Voxtral Realtime released?
Voxtral Realtime was released by Mistral AI on February 4, 2026.
How long a context window does Voxtral Realtime have?
Voxtral Realtime supports a context window of up to 128K tokens.

Official links

Sources

Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.

Release date and facts last verified November 11, 2026.

Related models

Model Released Parameters Context License Smallest variant needs
Qwen3-ASR Alibaba Qwen 918M – 2.35B – Apache 2.0 (Permissive) 2.12 GB Laptop or phone
Whisper large-v3-turbo OpenAI 810M – 1.54B – MIT (Permissive) 6 GB Laptop or phone
Nemotron 3.5 ASR NVIDIA 638M – OpenMDW-1.1 (Permissive) 2.93 GB Laptop or phone
Voxtral TTS Mistral AI 4.09B – CC BY-NC (Non-commercial) 9.2 GB Laptop or phone