Advertisement

Abu Dhabi developers introduce Jais 3 Arabic AI suite

Abu Dhabi researchers and technology company Inception42 have released Jais 3, a family of Arabic-focused artificial intelligence models designed to recognise speech, generate writing and complete tasks using software tools.

The Institute of Foundation Models at Mohamed bin Zayed University of Artificial Intelligence unveiled the system with Inception42 on Thursday, expanding the Jais programme beyond conversational responses into speech recognition and automated workflows. The release comprises six open language models, ranging from 900 million to 375 billion parameters, alongside a dedicated speech recognition model called Jais Speech.

The developers said the technology understands Modern Standard Arabic and regional dialects, including varieties spoken in the Gulf and North Africa. It is designed to maintain the appropriate dialect and level of formality across lengthy documents, while interpreting detailed instructions and producing written material suited to different users.

Jais Speech supports dialects spoken across 19 Arab countries and can transcribe conversations containing switches between Arabic, English and French. It preserves foreign-language words in their original scripts, identifies dialects and can insert Arabic vowel marks into transcripts. The speech system can also be combined with the language models to turn spoken instructions into actions.

The organisations assessed Jais 3 against competing systems using 24 Arabic benchmarks. They said the largest model matched leading open alternatives for Arabic comprehension and performed better on tests involving autonomous task execution. The speech model achieved the lowest average word and character error rates among systems examined across six public datasets, including commercial services.

Those findings reflect evaluations disclosed by the developers rather than a comprehensive independent comparison of every available model. Performance in individual organisations will also depend on the quality of their data, deployment arrangements and the tasks assigned to the system.

Among smaller models, the 3.7-billion-parameter version surpassed competing open models with up to five billion parameters on the developers' Arabic and agent-based assessments. The seven-billion- and 32-billion-parameter versions also achieved the highest average Arabic scores within their respective comparison groups.

At seven billion parameters, the reported average Arabic evaluation score increased from 56.4 for the underlying K2 Horizon model to 79.6 for Jais 3. The developers presented that improvement as evidence that an existing foundation model can be adapted for specialised linguistic requirements.

Haonan Li, a senior research scientist at the institute, said Arabic systems must account for differences between regions and generations, including speakers who change languages midway through a sentence. Such variations can complicate transcription, interpretation and the execution of spoken requests.

Jais 3 also supports tasks requiring several steps. When connected to external tools, it can search for information, edit documents and spreadsheets, and produce presentations or websites in Arabic. During a demonstration, the system researched forecasts concerning the 2026–2027 El Niño period and assembled a four-slide Arabic presentation using published information.

Larry Murray, senior vice-president of applied science at Inception42, said improved reasoning, instruction following and tool use would help businesses develop automated workflows. The availability of models with different computing requirements would allow organisations to select versions suited to their infrastructure and budgets.

The models are based on K2 Horizon, an open foundation model family released by the institute in September. Jais 3 represents its first adaptation to another language, with the developers applying additional Arabic linguistic and cultural training to the underlying architecture.

All six language models support a context window of 512,000 tokens, allowing them to process substantial amounts of text within a single request. The developers describe this as the longest context window offered by an Arabic-native model, although practical results may vary with document complexity.

Richard Morton, the institute's vice-president and executive director, said the release was intended to address Arabic as it is spoken, written and used in professional work. Inception42 chief executive Ashish Koshy said building an Arabic language model required more than translating material from other languages.

The models already underpin the Jais chat and mobile applications. A separate platform for agent-based workflows is planned, although the developers have not specified a release date or commercial pricing arrangements.

The project was developed with Cerebras Systems. The language models and Jais Speech are available through Hugging Face, with deployment support for vLLM and SGLang. The institute is also publishing its evaluation benchmarks and associated code, while providing five training checkpoints for each model to support further research and adaptation.
Previous Post Next Post

Advertisement

Advertisement

نموذج الاتصال