Voxtral Mini 4B Realtime Arabic

Not yet independently verified. The release date of 8 October 2026 rests only on the creation of the Hugging Face repository; Mistral AI published no announcement that we found, so the date may be earlier than the public release. The error rates are Mistral's own, from the model card, so no score is stored. No independent report was found. We will update this when it can be confirmed, and remove this note.

Voxtral Mini 4B Realtime Arabic is a streaming speech-to-text model from Mistral AI for Arabic dialects and Modern Standard Arabic. Its Hugging Face repository was created on 8 October 2026. It was built to cope with real-world audio in which speakers switch between Arabic and other languages.

What Voxtral Mini 4B Realtime Arabic does

The model transcribes 16 kHz audio as it arrives, and the delay between speech and text can be configured, which suits live captions and voice interfaces. It has about 4.4 billion parameters in BF16 and pairs a causal audio encoder with a language decoder. It was fine-tuned from Voxtral-Mini-4B-Realtime-2602, itself built on a 3B Ministral base.

Mistral reports an average character error rate of 8.82 per cent across seven Arabic benchmarks at a 480 millisecond delay, against 7.91 per cent for its Voxtral Transcribe Arabic. These are the maker's own figures, and the card warns that scores change with the normalisation method. Two new benchmarks, ISMA and Darija in the Wild, were developed with Morocco's digital transition ministry under a partnership signed in January 2026.

Who it is for

It suits developers who need live Arabic transcription they can host themselves, including for dialect-heavy and mixed-language speech. For other languages the card points to the original Voxtral Mini 4B Realtime. The model can be served with vLLM, streaming audio over a WebSocket at /v1/realtime, or run with Transformers version 5.2.0 or later.

Key facts

• Maker: Mistral AI, with Morocco's Ministère de la Transition Numérique et de la Réforme Administrative
• Repository created: 8 October 2026
• Licence: Apache 2.0; the card also says the model must not be used in ways that infringe third-party rights
• Pricing: open weights; no price on the model card
• Languages: Arabic dialects and Modern Standard Arabic
• Weights: Hugging Face, mistralai/Voxtral-Mini-4B-Realtime-Arabic