[Models](/models)/[Speech & Audio](/models?category=speech)/mistralai/Voxtral-Mini-4B-Realtime-2602

HFTranscriptionapache-2.0

# Voxtral-Mini-4B-Realtime-2602

by mistralai

Open-source realtime streaming speech-to-text with sub-500ms latency across 13 languages

2.4Mdl/month

952likes

4.4Bparams

[HuggingFace](https://huggingface.co/mistralai/Voxtral-Mini-4B-Realtime-2602)[Run it on your own data, free](https://studio.mixpeek.com/namespaces/create?mode=managed&utm_source=homepage&utm_medium=model_page&utm_campaign=mvs_onramp)

Identifiers

Model ID

`mistralai/Voxtral-Mini-4B-Realtime-2602`

## Deploy Voxtral-Mini-4B-Realtime-2602

Single-tenant

Mixpeek has no managed extractor for this model. On a single-tenant deployment you upload the weights and a custom plugin serves them next to the rest of your pipeline.

[Read the customer-hosted guide](https://mixpeek.com/docs/customer-hosted/overview) [Talk to us about single-tenant](/demo)

## Overview

Voxtral Mini 4B Realtime is among the first open-source speech models to achieve offline-comparable accuracy with sub-500ms latency. Its natively streaming architecture pairs a causal audio encoder (~0.6B params) with a Ministral-3-based LLM decoder (~3.4B params), both using sliding window attention for constant-memory streaming inference.

On Mixpeek, Voxtral powers realtime and near-realtime transcription of audio and video content across 13 languages, with configurable latency from 240ms to 2.4s to balance speed against accuracy for live subtitling or batch processing.

## Architecture

Two-component streaming architecture: (1) causal transformer audio encoder (0.6B params, 32 layers, causal attention) and (2) Ministral-3-based LLM decoder (3.4B params, 26 layers). Both use sliding window attention for streaming. Configurable transcription delay from 240ms to 2.4s.

## How it runs

.dataflow-R15f6b .df-packets-static{display:none} @media (prefers-reduced-motion: reduce){ .dataflow-R15f6b .df-packets-anim{display:none} .dataflow-R15f6b .df-packets-static{display:inline} .dataflow-R15f6b .df-toggle{display:none} }

Pause

Inference INPUT MODEL OUTPUT Audio object Voxtral-Mini-4B-Realtime-2602 mp.inference Transcript vector Vector store MVS Voxtral-Mini-4B-Realtime-2602 → embeddings, indexed for search

Voxtral-Mini-4B-Realtime-2602 takes audio, and Mixpeek indexes what it emits.

## Mixpeek SDK Integration

IngestRetrieve

// No extractor parameter takes a Hugging Face model id (checked against
// GET /v1/discovery/extractors, which returns 13), so Voxtral-Mini-4B-Realtime-2602 runs
// on your side and the output is upserted through POST
// /v1/namespaces/{namespace\_id}/documents/upsert. On Enterprise the other
// path is to upload the weights instead: POST /v1/namespaces/{id}/models
// accepts the huggingface format and a custom plugin loads them.
const res = await fetch(
  "https://api.mixpeek.com/v1/namespaces/ns\_your\_namespace/documents/upsert",
  {
    method: "POST",
    headers: {
      Authorization: "Bearer API\_KEY",
      "Content-Type": "application/json",
    },
    body: JSON.stringify({
      collection\_id: "col\_your\_collection",
      documents: \[
        {
          document\_id: "asset-00412",
          // The model produces text, so it lands in payload. Give the
          // collection a text vector index and embed that text to make it
          // searchable rather than only filterable.
          payload: { extracted\_text: modelOutput, source\_key: "archive/2026/asset-00412" },
          vectors: { "text-embedding": embeddingOfModelOutput },
        },
      \],
    }),
  },
);

// Managed alternative, if this exact model is not the requirement:
// universal\_extractor@v1 runs google/gemini-embedding-2
// (3072-d) over a bucket, with no inference of your own.

## Capabilities

*   Realtime streaming transcription with <500ms latency
*   13 language support including English, Spanish, French, German
*   Configurable latency/accuracy tradeoff (240ms-2.4s delay)
*   Natively streaming architecture (no chunking workarounds)
*   Apache 2.0 open-source

## Use Cases on Mixpeek

Live subtitling and closed captioning for video streams

Voice assistant transcription with low-latency requirements

Multilingual meeting transcription with realtime output

## Benchmarks

Dataset

Metric

Score

Source

FLEURS (13 languages, 480ms)

Average WER

8.72%

Mistral AI, Feb 2026: Voxtral Realtime paper

FLEURS English (480ms)

WER

4.90%

Mistral AI, Feb 2026: Voxtral Realtime paper

FLEURS (13 languages, 2.4s)

Average WER

6.73%

Mistral AI, Feb 2026: Voxtral Realtime paper

## Performance

Input SizeStreaming audio (16kHz)

GPU Latency240ms-2.4s configurable delay (A100)

GPU ThroughputRealtime factor >1x (streaming)

GPU Memory~8.5 GB

## Common Pipeline Companions

[

BAAI/bge-m3

Embed transcripts for multilingual text search

](/model/BAAI/bge-m3)[

openai/whisper-large-v3

Compare streaming vs offline transcription quality

](/model/openai/whisper-large-v3)

## Explore on Mixpeek

[

More Speech & Audio models

Compare alternatives in this category

](/models?category=speech)[

Best Audio Processing Tools

Hand-picked tools & platforms compared

](/curated-lists/best-audio-processing-tools)[

ASR Decoding: Beam Search & LM Fusion

Deep-dive technical guide

](/guides/asr-decoding-beam-search-lm-fusion)[

Feature Extractors

See how Mixpeek runs models as extractors

](/extractors)[

Mixpeek Vector Store

Store & search embeddings at scale

](/mvs)[

Pricing

Usage-based pricing for pipelines

](/pricing)[

All Curated Lists

Compare models, APIs & infrastructure

](/curated-lists)

### Specification

FrameworkHF

Organizationmistralai

FeatureTranscription

Outputtext + timestamps

Modalitiesvideo, audio

RetrieverTranscript Search

Parameters4.4B

Licenseapache-2.0

Downloads/mo2.4M

Likes952

[

### Research Paper

Voxtral Realtime

arxiv.org](https://arxiv.org/abs/2602.11298)

### Build a pipeline with Voxtral-Mini-4B-Realtime-2602

Add this model to a processing pipeline alongside other extractors. Combine with retrieval stages for end-to-end search.

[Run it on your own data, free](https://studio.mixpeek.com/namespaces/create?mode=managed&utm_source=homepage&utm_medium=model_page&utm_campaign=mvs_onramp)

### Alternative Models

[

openai/whisper-large-v3

Transcription

](/model/openai/whisper-large-v3)[

facebook/wav2vec2-large-960h

Transcription

](/model/facebook/wav2vec2-large-960h)[

distil-whisper/distil-large-v3

Transcription

](/model/distil-whisper/distil-large-v3)[

nvidia/parakeet-tdt-0.6b-v3

Transcription

](/model/nvidia/parakeet-tdt-0.6b-v3)

### Related in Speech & Audio

[

pyannote/speaker-diarization-3.1

Speaker Diarization

](/model/pyannote/speaker-diarization-3.1)[

pyannote/speaker-diarization-community-1

Speaker Diarization

](/model/pyannote/speaker-diarization-community-1)[

nvidia/diar\_sortformer\_4spk-v1

Speaker Diarization

](/model/nvidia/diar_sortformer_4spk-v1)[

pyannote/wespeaker-voxceleb-resnet34-LM

Speaker Diarization

](/model/pyannote/wespeaker-voxceleb-resnet34-LM)