AI Expertise
Open Models / Voxtral Mini 4B Realtime
V

Voxtral Mini 4B Realtime

Visit website

Description

Voxtral Mini 4B Realtime 2602 is an open model that transcribes speech as it is spoken. It has 4 billion parameters and a natively streaming architecture with a custom causal audio encoder. Per the model card, the delay is configurable from 240 ms to 2.4 s, and at 480 ms the authors say it matches leading offline open models and realtime APIs. It supports 13 languages: English, French, Spanish, German, Russian, Chinese, Japanese, Italian, Portuguese, Dutch, Arabic, Hindi and Korean. It is meant for voice assistants and live subtitles, and it is optimized for on-device deployment. The weights are in BF16 and, according to the card, run on a single GPU with 16 GB of memory or more; the recommended way to serve it is vLLM, with a default maximum length of about 3 hours of audio. On the Open ASR Leaderboard (English short-form speech, nine test sets, checked 2026-10-20) it has an average word error rate of 9.01%; this is a general benchmark, not a live-streaming one. Released 2026-01-21. The weights are public without any acceptance step under the Apache-2.0 license, which allows commercial use. Date of check: 2026-10-20.

Pros and cons

Pros

  • Real-time transcription with a delay from 240 ms to 2.4 s, per the model card
  • 13 languages
  • Runs on a single GPU with 16 GB of memory or more
  • Apache-2.0 license, weights open without an acceptance step
  • Up to about 3 hours of audio in one session by default

Cons

  • The card does not mention timestamps or speaker separation
  • Needs a 16 GB GPU; the card recommends vLLM
  • Highest average error rate among the models in our list on the Open ASR Leaderboard (a non-streaming benchmark)
  • Only 13 languages

Key facts

License
Apache-2.0
Commercial use
Allowed
Hardware requirements
A single GPU with 16 GB of memory or more
Developer
Mistral AI
Interface languages
English
Checked on
Oct 20, 2026

Editorial score

How strongly the AI Expertise editors recommend this open model for each of its tasks right now, on a 0–10 scale. The score combines output quality, capabilities, ease of launch, license openness and freshness.

Transcription

6.8 / 10
Output quality 7.4
Low High
Features 2.9
Basic Rich
Ease of launch 7.0
Hard Easy
License openness 10.0
Restricted Open
Freshness 6.0
Outdated Fresh

News about Voxtral Mini 4B Realtime

  • Mistral AI has launched Voxtral 3, a real-time voice-agent stack combining listening, reasoning, tool use and speech in one streaming pipeline. It supports transcription in 50+ languages, expressive speech in 35+ languages and 50 dialects, and can run on your infrastructure or EU-hosted APIs. Transcription starts at $0.0035/min. #AI #VoiceAgents #MistralAI #Voxtral
  • Mistral AI launches Voxtral Transcribe 2, a new speech recognition model family starting at $0.003/min — undercutting competitors while reportedly outperforming GPT-4o mini, Gemini 2.5 Flash, and Deepgram Nova. Two variants: Voxtral Mini Transcribe V2 for long audio (up to 3 hrs) and Voxtral Realtime for sub-200ms latency apps. Both support 13 languages, speaker diarization, and word-level timestamps. Realtime is open-weights (Apache 2.0) on Hugging Face; Mini is API-only. First gen debuted July 2025. #MistralAI #Voxtral #SpeechRecognition #AI #OpenSource
  • Mistral AI launches Voxtral Transcribe 2 at $0.003/min, undercutting competitors. Claims superior accuracy over GPT-4o mini, Gemini 2.5 Flash, Deepgram Nova. Two models: Mini for large files, Realtime (<200ms latency) for voice assistants/live captioning. Both support 13 languages with speaker diarization, word timestamps, 3-hour recordings. Realtime open-weight (Apache 2.0) on Hugging Face; Mini via API only. #MistralAI #Voxtral #SpeechRecognition #AI #OpenSource
All news about Voxtral Mini 4B Realtime
Report an error