Why We Need Arabic Language Models – Research Highlights

In the
global race to develop generative AI models, attention tends to center on major
companies and research institutes in the West and China. Flagship models, including
OpenAI’s GPT-4 and Google’s Gemini, are trained on vast amounts of data,
predominantly in English and other Western languages, and therefore tend to
reflect the cultural assumptions and values of the context in which they were
developed.

The
growing reliance on language models that do not necessarily reflect the richness and diversity of Arabic language poses a
significant challenge. It’s not simply a matter of technical preference, but
one that raises questions of cultural sovereignty, technological independence,
and national identity.

These
widely used models such as ChatGPT have the potential to shape perceptions and
ideas. When trained on data from different cultural contexts, these models can
generate responses that sideline core Arab values or remain vague on critical
issues.

Clear examples emerge when global language models
address culturally sensitive issues, such as social relationships or political
debates. They often adopt ambiguous positions that overlook the Arab cultural
context, creating a gap between these digital tools and the values and lived experiences
of Arab users.

The lack of robust and competitive
Arabic language models forces researchers and developers across the region to
rely on tools that
fail to capture the linguistic complexity of Arabic,
its dialects, or cultural contexts. This dependence constrains the ability to
design AI applications and services tailored to local needs, while also weakens
the Arab world’s contribution to global AI advancement. In many ways,
language models serve as a mirror of our research and innovation capacity.

In response to this challenge, promising
initiatives have emerged across the Arab world. For instance, the UAE’s ‘Jais,’
Saudi Arabia’s ‘ALLaM,’ and Qatar’s ‘Fanar,’ which was developed by the Qatar
Computing Research Institute (QCRI) at Hamad Bin Khalifa University (HBKU) in
collaboration with government partners. These initiatives are part of broader
strategic efforts to localize technology, safeguard cultural identity, and build
technological self-reliance.

Developing
such models, however, comes with significant challenges. One of the most
persistent is the scarcity of high-quality Arabic content online compared with
English. While Fanar was trained on more than half a trillion Arabic words,
this remains modest when compared to global models trained on trillions of
tokens. The quality of available Arabic data also varies widely, due to
accuracy issues, linguistic style, and considerable diversity between Modern
Standard Arabic and regional dialects, making data collection and
representation more complex.

Another major challenge is the high cost
of training large language models. For example, training a 7-billion-parameter
model on a trillion words requires more than 220 H100 GPUs running continuously
for over a month, and these resources are often beyond the reach of most
research institutions in the Arab world. This inspired the Fanar team to focus
on developing smaller models with seven and nine billion parameters,
prioritizing improvements in data quality and optimization techniques to deliver the best possible performance with
the resources available.

Addressing
the challenges of cultural and technological dependency requires collaboration
across multiple sectors. Academic and research institutions need to invest in
Arabic language processing and build international partnerships to maximize
resources and expertise. Governments and policymakers, in return, should
provide sustained funding; support data infrastructure; promote policies that
facilitate the collection and organization of high-quality Arabic datasets; and
foster collaboration between the public and private sectors, as it is essential
for building a supportive ecosystem for technological innovation.

Startups and developers in the region
also have a role to play. Both should adopt Arabic language models to build applications
that respond to local needs, from AI personalized education platforms to
voice assistants in regional dialects. Cultural, educational, and media
institutions, meanwhile, can contribute by generating diverse, high-quality
Arabic digital content that can
be used to train these models.

Building
robust Arabic language models is not a technological luxury, but a strategic necessity
to ensure that the Arab world has a voice in shaping the future of AI. While
significant progress has been made, the path ahead requires sustained
investment and collective effort from stakeholders across the region.

_{This is a translation of the Arabic article published on 3rd
August 2025}