#UAE #Arabic - Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) Institute of Foundation Models and G42’s sovereign agentic AI company Inception42 have launched Jais 3, a fleet of six open Arabic language models plus Jais Speech, a speech recognition model. The models range from 900 million to 375 billion parameters and are built on K2 Horizon, IFM’s open-source foundation model fleet. They handle Arabic dialects, writing and multi-step agentic tasks. Both are available now on the AI community site Hugging Face.
SO WHAT? - Jais 3 gives Arabic developers a choice of sizes, from a model small enough for a phone to a 375-billion-parameter flagship model, all with the same agent capabilities. Speech is the practical addition, and it includes ‘code-switching’, which supports speakers who switch between Arabic and English or French mid-sentence. The scores come from the developers’ own evaluations, though they’re publishing the 24 benchmarks used. The 7-billion-parameter model’s jump from 56.4 to 79.6 on those tests shows how far adapting an existing open model to a new language can go.
KEY POINTS:
Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) and its Institute of Foundation Models (IFM), together with Inception42 have released Jais 3. The model family was developed in collaboration with Cerebras Systems.
The fleet has six open language models, from 900 million to 375 billion parameters, plus Jais Speech for Arabic and its dialects. All are built on K2 Horizon, IFM’s open-source foundation model fleet released in September 2026.
Training covered Modern Standard Arabic and regional dialects, from Emirati and Gulf to North African Arabic, along with cultural knowledge. All models share one training recipe, agent capabilities and a 512,000-token context window, which the developers say is the longest of any Arabic-native model.
The developers evaluated the models on an audited suite of 24 Arabic benchmarks, which they’re releasing publicly. They say the largest model leads open models on agentic tasks and matches the strongest on Arabic understanding.
The 3.7-billion-parameter model outperformed every open model of up to 5 billion parameters on Arabic and agentic tasks. The 7-billion and 32-billion-parameter models had the highest average Arabic scores among open models at their sizes.
Jais Speech transcribes Modern Standard Arabic and dialects from 19 Arab countries. It keeps English or French words in their original script, identifies the dialect and can add Arabic vowel marks. It recorded the lowest average word and character error rates of any model compared, including commercial services.
Paired with the language models, Jais Speech lets Jais 3 take spoken instructions in Arabic or English and act on them. At 7 billion parameters, the average Arabic score rose from 56.4 for the K2 Horizon base model to 79.6 for Jais 3.
Each model ships with five training checkpoints for researchers. UAE and international government entities and enterprises can adapt the models and choose where to deploy them, in line with sovereign AI requirements. They support vLLM and SGLang.
[Written and edited with the assistance of AI]
Source: MBZUAI, IFM, Inception42
LINKS
Jais landing page (IFM)
Jais 3 Collection (Hugging Face)
Jais 3 chat demo (demo site)
Read more about Arabic model development:
Egyptian startup releases three open-source Arabic models (Middle East AI News)
TII launches Arabic Falcon model for Emirati dialect (Middle East AI News)
Morocco, Mistral partnership releases Darija AI (Middle East AI News)
TokenAI open-sources Horus Taleeq Arabic model (Middle East AI News)
HUMAIN unveils 428B Arabic model with MiniMax (Middle East AI News)