The Institute of Foundation Models (IFM) at Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) and Inception42 have introduced Jais 3, a family of open Arabic AI models designed to handle regional dialects, speech recognition, document creation, and multi-step agent workflows. The release combines six language models ranging from 900 million to 375 billion parameters with Jais Speech, an Arabic speech recognition model intended to support real-world conversations, including switches between Arabic, English, and French.
Arabic-native AI moves beyond translation
Jais 3 is designed around the way Arabic is actually spoken and written, including Modern Standard Arabic, regional dialects from the Gulf to North Africa, and conversations in which speakers switch between Arabic and English or French. The models also target culturally specific knowledge, including Arabic poetry, cuisine, and Islamic subjects.
That focus addresses a challenge for multilingual AI systems: broad language coverage does not necessarily translate into consistent performance across regional speech, cultural references, or detailed writing instructions. For organizations operating across the Middle East and North Africa, those differences can affect whether an AI application is useful in customer service, document processing, internal research, and administrative workflows.
The model family spans six sizes, from a 900-million-parameter version intended for constrained hardware to a 375-billion-parameter flagship. IFM says the models share a common training approach and support agentic capabilities, allowing developers to choose a configuration based on available computing resources and deployment requirements.
Speech recognition and agent workflows
Jais Speech adds Arabic automatic speech recognition to the model family. It is designed to transcribe Modern Standard Arabic and dialects spoken across 19 Arab countries, including speech recorded in noisy environments. It can preserve English and French words in their original scripts, identify the dialect being spoken, and add Arabic vowel marks to transcripts.
In evaluations reported by IFM and Inception42, Jais Speech recorded the lowest average word and character error rates among the models they compared across six public test sets. The organizations also report strong results for Jais 3 on a suite of 24 Arabic benchmarks. These evaluations were conducted by the developers, so independent replication would help establish how well the results generalize to different applications.
The broader goal is to combine language understanding with the ability to complete tasks. Connected to software tools, Jais 3 can plan multistep workflows, search the web, and create or edit Arabic documents, spreadsheets, presentations, and websites.
In one demonstration described by the developers, a model researched the 2026–2027 El Niño forecast and produced a four-slide Arabic presentation using published forecast data. The example illustrates a workflow that moves from information retrieval to a finished document rather than stopping at a conversational response.
Open models and regional AI deployment
Jais 3 builds on K2 Horizon, IFM’s open model family released in September 2026. At the seven-billion-parameter size, the organizations report that the average Arabic evaluation score rose from 56.4 for the K2 Horizon base model to 79.6 for Jais 3. The result suggests that adapting an existing foundation model to a specific language can produce substantial gains without training an entirely new model from scratch.
IFM is also releasing its audited Arabic benchmarks and evaluation code, along with five training checkpoints for each model. This gives researchers more opportunities to inspect model development and test performance, although the practical openness of a model also depends on the completeness and licensing of its weights, data, and training materials.
The models are intended for organizations that need flexibility over computing costs and deployment locations, including governments pursuing sovereign AI strategies. Smaller versions may support local or edge deployment, while larger models target more demanding workloads.
Jais 3 and Jais Speech are available through Hugging Face, with deployment support for vLLM and SGLang. The Jais chat and mobile apps also use the models, while a dedicated agentic platform for everyday workflows is planned.
The launch highlights a broader direction in AI infrastructure: building models for languages and local contexts that general-purpose systems may not consistently handle. Jais 3’s long-term value will depend on independent testing, deployment costs, and performance in real-world workflows. For Arabic-speaking organizations, the combination of speech, writing, and tool use offers a route to AI systems designed around their language rather than treating it as an afterthought.
Market Landscape
Arabic-native AI is an important area of development as enterprises and governments seek language models that work across regional dialects, cultural contexts, and local data requirements. Jais 3 enters a market where multilingual foundation models compete with specialized systems built for particular languages and workflows.
Its main differentiators are the combination of six model sizes, Arabic speech recognition, agentic capabilities, and access to evaluation benchmarks. The reported improvement from a 56.4 to 79.6 average Arabic score at seven billion parameters also highlights the potential of adapting an existing open foundation model to a language-specific context.
The competitive test will be whether these advantages hold up under independent evaluation and production workloads. Organizations will also need to assess hardware requirements, deployment costs, licensing, and the accuracy of outputs in their own dialects and domains.
Top Insights
- Jais 3 combines six open Arabic language models with speech recognition, covering model sizes from 900 million to 375 billion parameters.
- Jais Speech supports transcription across dialects spoken in 19 Arab countries, including conversations that switch between Arabic, English, and French.
- IFM and Inception42 report improvements on 24 Arabic benchmarks, although independent evaluations are needed to validate performance claims.
- Tool integration enables Jais 3 to search the web, plan multistep tasks, and create Arabic documents, spreadsheets, presentations, and websites.
- Public benchmarks, training checkpoints, and flexible deployment options could help developers and governments evaluate and adapt Arabic AI for local requirements.
Power Tomorrow’s Intelligence — Build It with TechEdgeAI
