Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) and Inception42 released six open Arabic-language models on Oct. 15, pairing them with speech recognition for dialects used in 19 Arab countries. The models give developers tools for Arabic-language applications across a range of computing setups.
MBZUAI’s Institute of Foundation Models (IFM) and Inception42 developed the text models with Cerebras Systems. The release also includes Jais Speech. The text models span 900 million to 375 billion parameters. MBZUAI says they are available on Hugging Face, while IFM says the models and checkpoints use the Apache 2.0 license. MBZUAI says Arabic has more than 400 million speakers.
MBZUAI says Jais Speech transcribes Modern Standard Arabic and dialects used across the 19 countries. The system can identify a speaker’s dialect, add vowel signs to Arabic text and retain English or French words in their original script when speakers switch languages. The developers say Jais 3 can follow spoken instructions and create documents and presentations.
IFM’s deployment guidance assigns the 900-million-parameter model to phones and edge settings, with larger versions intended for machines with greater computing capacity. MBZUAI says governments and businesses can adapt the models for uses including sovereign AI. Inception42 says organizations can choose among model sizes to match their computing budgets and deployment requirements.
IFM says about 59% of its Arabic training tokens were synthetic or machine-translated. It reports that about 36% came from dialect-focused sources, including Arabizi, parallel dialect material and speech transcripts.
In its Jais-ASR model card, IFM reports average error rates of 22.85% at the word level and 8.37% at the character level across six public test sets, which it says were the lowest in its comparison. Separately, Data Phoenix reported that the developers’ benchmark results had not been independently replicated.
The National reported limited open data for Gulf dialects and less even Moroccan dialect coverage outside Darija. Haonan Li, a senior research scientist at IFM, told the newspaper that ensuring “broad and balanced representation across regional dialects” was a challenge.
