United Arab Emirates AI Models: A Guide

The UAE has become one of the most active AI builders outside the US and China. Backed by state investment and anchored in Abu Dhabi, institutions such as the Technology Innovation Institute (TII), Mohamed bin Zayed University of Artificial Intelligence (MBZUAI), G42 and its subsidiaries Inception and Core42 have released models that compete globally, many of them open source. This article lists the main UAE-developed models and breaks down what each does best. Model families evolve quickly, so check the official repositories for the latest versions.

Overview Table

Model	Developer	Type	Focus	First Release
Falcon family	TII	Open-weight LLMs (Transformer, Mamba, hybrid)	General purpose, efficiency, multilingual	2023
Jais family	Inception (G42), MBZUAI, Cerebras	Bilingual LLMs	Arabic and English	2023
K2 Think	MBZUAI and G42	Reasoning model (32B)	Math, code, science reasoning	2025
Nanda	Inception, MBZUAI, Cerebras	Bilingual LLM (87B)	Hindi and English	2024
Sherkala	Inception, MBZUAI, Cerebras	Bilingual LLM (8B)	Kazakh and English	2025
BiMediX	MBZUAI	Medical LLM	Bilingual healthcare Q&A	2024
PAN	MBZUAI	World model	Simulating and predicting environments	2025

1. Falcon (TII)

Falcon is the UAE’s best-known model line. TII released Falcon 40B and 180B in 2023, and 180B briefly topped open-model leaderboards. Later releases include Falcon 2, Falcon Mamba 7B, Falcon 3, Falcon Arabic and the hybrid Falcon-H1 series.

Features

• Sizes from sub-1B models up to 180B parameters
• Architectures spanning pure Transformer, state-space (Mamba) and hybrid Transformer-Mamba designs
• Falcon 2 added vision-to-language capability
• Falcon Arabic targets native Arabic dialect handling
• Permissive, open-access licensing for most releases

Strengths

• Strong performance per parameter, so small models can run on laptops and edge devices
• Efficient long-context processing thanks to Mamba-based designs
• Wide size range lets teams match the model to their hardware budget
• Large developer community and extensive fine-tunes

2. Jais (Inception, MBZUAI, Cerebras)

Jais launched in 2023 as one of the first large language models built from the ground up for Arabic. The family later expanded to many sizes, trained from scratch and adapted from Llama, and a larger second-generation Arabic model followed.

Features

• Bilingual Arabic-English training on large Arabic corpora
• Parameter sizes ranging from hundreds of millions to 70B
• Instruction-tuned chat variants
• Built on Cerebras supercomputing infrastructure
• Open weights available for research and commercial use

Strengths

• Understanding of Arabic grammar, culture and regional context that general models often miss
• Competitive English ability despite the Arabic focus
• Suited to government, banking and education use cases in the Arab world
• Strong choice for Arabic chatbots and document processing

3. K2 Think (MBZUAI and G42)

K2 Think, announced in 2025, is a 32B-parameter reasoning system built on an open base model and tuned with long chain-of-thought training, reinforcement learning and test-time scaling techniques.

Features

• Step-by-step reasoning for math, coding and scientific problems
• Open release of weights, training approach and evaluation details
• Designed for fast inference on specialised hardware

Strengths

• Reasoning scores that rival much larger proprietary models
• Parameter efficiency, which lowers deployment cost
• Transparency, which helps researchers reproduce and extend the work

4. Nanda (Inception, MBZUAI, Cerebras)

Nanda is an 87B-parameter model aimed at Hindi speakers, built from a Llama base and extended with extensive Hindi data. It shows the UAE’s ambition to serve languages beyond Arabic.

Features

• Hindi-English bilingual capability
• Support for Hinglish (mixed Hindi and English) usage patterns
• Trained with a large Hindi-focused corpus

Strengths

• Better Hindi fluency than most general-purpose open models at release
• Useful for South Asian customer support, media and education in the Gulf
• Large expat population in the UAE makes it directly relevant locally

5. Sherkala (Inception, MBZUAI, Cerebras)

Sherkala is an 8B-parameter model for Kazakh, extending the same bilingual approach to Central Asia.

Features

• Kazakh-English bilingual support
• Compact size suitable for modest hardware
• Open release for research and application building

Strengths

• Addresses a language with limited AI coverage
• Light enough for on-premise deployment
• Demonstrates a repeatable recipe for low-resource languages

6. BiMediX (MBZUAI)

BiMediX is a bilingual medical language model that answers healthcare questions in Arabic and English.

Features

• Fine-tuned on medical instruction data in both languages
• Supports multi-turn medical conversations and multiple-choice exam-style questions
• Evaluated on a dedicated Arabic-English medical benchmark

Strengths

• Brings clinical knowledge to Arabic speakers, who are underserved by medical AI
• Useful for research into healthcare assistants and medical education tools
• Domain specialisation improves accuracy over general chat models on medical questions

Note: It is a research tool and should not replace professional medical advice.

7. PAN (MBZUAI)

PAN is a general-purpose world model from MBZUAI’s Institute of Foundation Models. Rather than just generating text, it simulates how environments change over time in response to actions.

Features

• Combines language-based reasoning with video generation to predict future states
• Supports long-horizon, action-conditioned simulation
• Aimed at planning, robotics and interactive environments

Strengths

• Moves beyond chatbots toward models that understand cause and effect
• Useful for training agents in simulated settings
• Positions the UAE in a frontier research area

What Ties These Models Together

Three themes stand out across the UAE’s portfolio. First, openness: most models ship with open weights, which encourages adoption and scrutiny. Second, language inclusion: Jais, Nanda and Sherkala prove that regional languages can get first-class models. Third, efficiency: Falcon and K2 Think show that smaller, well-trained models can match larger ones.

The UAE’s AI model ecosystem is broader than many expect. Falcon offers general-purpose efficiency, Jais anchors Arabic NLP, K2 Think pushes open reasoning, and Nanda, Sherkala, BiMediX and PAN extend the reach into new languages, medicine and world simulation. Developers choosing among them should weigh language needs, hardware limits and licensing, and verify current versions and benchmarks on the official TII, MBZUAI and Inception channels.