MBZUAI introduces six open Arabic AI models, alongside regional speech recognition
The developers describe document creation and multi-step software tasks across the lineup, plus transcription covering dialects spoken across 19 Arab countries.
Loading page…
The developers describe document creation and multi-step software tasks across the lineup, plus transcription covering dialects spoken across 19 Arab countries.
Listen to this story
MBZUAI’s Institute of Foundation Models and Inception42, with Cerebras Systems, announced Jais 3 on October 15, 2026: six open Arabic-language models built on K2 Horizon, plus companion Jais Speech. The 900-million-to-375-billion-parameter range is intended to span constrained devices through data-center-scale deployments, though those settings are positioning claims, not demonstrated hardware requirements. Developers also claim software-tool workflows across the family and speech coverage for Modern Standard Arabic and dialects across 19 Arab countries, making this a broader Arabic AI stack rather than a text-only launch.
Developers say all six Jais 3 models can create documents and presentations and perform multistep tasks using software tools; these capabilities remain announcement claims.
Jais Speech is designed to transcribe Modern Standard Arabic and dialects spoken across 19 Arab countries; users can give it spoken instructions in Arabic or English.
Jais 3 adapts K2 Horizon, so the announcement does not present it as a foundation model built from scratch.
Arabic AI that does work, rather than only answers questions, is the pitch behind Jais 3. MBZUAI’s Institute of Foundation Models and Inception42 announced the family on October 15, 2026, in collaboration with Cerebras Systems. The release pairs six open language models with speech recognition; its task capabilities remain developer claims.
The smallest model is positioned for constrained hardware, including phones and edge devices. The largest targets data centers or large GPU clusters. These are intended deployment settings, not tested hardware requirements. The announcement presents a range of options rather than a single model for every installation.
Jais 3 builds on K2 Horizon, an existing foundation-model base, with the new family focused on Arabic-language applications. It is an adaptation of that underlying technology, rather than a claim that the developers built an entirely separate foundation from scratch.
The six open Arabic-language models span 900 million to 375 billion parameters.
The developers describe support for Arabic text generation, document and presentation creation, and tasks requiring multiple steps through software tools. They place those tool-oriented capabilities across all six models, not only the flagship. The intended output extends beyond a conversational reply to files and completed sequences of actions in connected software.
The companion Jais Speech system has a different job: transcription, or turning speech into text. Its developers describe coverage of Modern Standard Arabic and dialects spoken across 19 Arab countries. That scope makes regional speech, rather than formal Arabic alone, part of the coordinated announcement.
According to the announcement, Jais Speech also supports spoken instructions in Arabic or English. The two claims address different parts of the experience: which speech the system is designed to transcribe, and which languages people can use to give it instructions.
Loading discussion...
Join the conversation
Explain when formal Arabic alone would—or would not—be enough.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.