The listed model name is cohere-transcribe-arabic-07-2026; it is a 2B-parameter, audio-in/text-out model licensed under Apache 2.0.
The documentation lists Arabic and English coverage, multidialectal support, and Arabic-English code-switching. Input files are limited to 25 MB. The model does not return timestamps and does not perform speaker diarization for multi-speaker recordings.
Cohere offers rate-limited API access for free experimentation. Production use without rate limits requires a dedicated Model Vault deployment, described as private cloud inference billed per hour-instance; the page does not state a price.
The source characterizes the model as state of the art and suited to varied acoustic conditions, bilingual dialogue, dialects, and enterprise vocabulary. It does not provide benchmark tables, evaluation methodology, a release date, or evidence allowing an independent comparison with other Arabic ASR models.