Voxtral Mini 4B Realtime Arabic

Voxtral Mini 4B Realtime Arabic is a speech-to-text model from Mistral AI that transcribes Arabic as the audio arrives. Its model card says it covers Arabic dialects and Modern Standard Arabic and is built to cope with speakers who switch between languages. It has about 4.4 billion parameters and pairs a causal audio encoder with a language decoder working on 16 kHz audio, with a transcription delay that can be set; the evaluations on the card used 480 ms. It is fine-tuned from Voxtral Mini 4B Realtime, and for other languages the card points users to that original model. The weights are open under the Apache 2.0 licence, with a clause barring use that infringes third-party rights, and the card shows serving through vLLM over a WebSocket and through Transformers. Mistral reports an average character error rate of 8.82 per cent across seven Arabic test sets, 0.91 points behind its Voxtral Transcribe Arabic result. These are the vendor's own figures, depend on the card's own text normalisation, and include two new sets developed with Morocco's national telecoms regulator. The card gives no audio length limit.

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.