HUMAIN Chat: Saudi Arabia's Arabic-language AI platform

The problem

In 2024, the generative AI tools widely available on the market offered limited and uneven support for Arabic. For Saudi Arabia, which has positioned AI as a central pillar of its Vision 2030 strategy, it was both a practical and a strategic problem. The Saudi Data and Artificial Intelligence Authority (SDAIA) identified two interlocking gaps.

The first gap was structural: Saudi Arabia had no domestically developed large language model, nothing built on Arabic datasets, by local teams, on national infrastructure. Any AI capability the team deployed would therefore sit on foreign platforms, under foreign data governance arrangements, with no meaningful Saudi control over how the underlying model was trained or updated. For a government with growing AI ambitions across public services, that dependency was becoming difficult to justify.

The second gap was linguistic and cultural: the dominant generative AI platforms were built primarily on English-language data, and their Arabic support typically relied on translation layers rather than native comprehension. The challenge with this is that Arabic is not a single, uniform language—it spans Modern Standard Arabic, classical forms, and a wide range of regional dialects, some of which are sufficiently distinct that speakers from different countries can struggle to understand each other. Existing models were not equipped to handle this range. Beyond weak dialect comprehension, cultural nuance, including alignment with Islamic values and Middle Eastern context, was largely absent from mainstream tools.

Saudi Arabia needed a model that could genuinely understand Arabic — in its full range — and one that they could own, develop, and govern themselves.

The solution

In response, SDAIA began developing a domestically built large language model focused specifically on Arabic. The programme was led through the National Centre for Artificial Intelligence and later brought together under HUMAIN.

The goal was clear: a national foundation model that could operate in Arabic as its primary language and run on infrastructure hosted within Saudi Arabia.

This work led to the development of ALLAM 34B, a large language model trained primarily on Arabic-language data. According to HUMAIN, the model was trained on more than 500 billion Arabic tokens, meaning individual units of text such as words, parts of words, and punctuation that the model learns patterns from. That makes it one of the largest Arabic-language training datasets assembled to date. The model was then refined with input from hundreds of domain experts and evaluators across sectors, with the aim of improving performance in Modern Standard Arabic as well as regional dialects and ensuring that outputs reflected relevant cultural and contextual references.

Building on this model, the team launched HUMAIN Chat — a chat-based AI platform hosted within Saudi Arabia. It was designed to make Arabic-first digital interaction practical for everyday use, including within public services. HUMAIN Chat allows users to interact with the ALLAM model through a chat interface to draft documents, summarise reports, translate between Arabic and English, and retrieve information in natural language.

The platform supports Arabic speech input across multiple dialects, so users can speak rather than type — and allows switching between Arabic and English within a single conversation, reflecting how many people in Saudi Arabia naturally move between languages. It also includes real-time web search and the ability to share conversations between users for handover or joint working

HUMAIN Chat is hosted within Saudi Arabia and operates in line with the country's Personal Data Protection Law (PDPL). It is designed not as a standalone consumer tool but as infrastructure that could be integrated into government services and other public-facing systems.

The impact

1. Linguistic inclusion at scale
The impact is both practical and symbolic.

Arabic speakers — including those using regional dialects — can interact with a nationally developed AI system without relying on translation layers or foreign-hosted platforms. That changes the tone of digital interaction.

For public services, language is not cosmetic. It shapes who feels confident engaging, who understands clearly, and who trusts the system responding. HUMAIN positions Arabic not as a secondary language in AI, but as a primary one.

2. Strengthening digital sovereignty and data localisation
According to HUMAIN’s official launch materials, the platform is hosted end-to-end within Saudi Arabia and positioned as compliant with Saudi Arabia’s Personal Data Protection Law (PDPL). This is significant in a policy context.

For governments, AI capability is no longer only a service innovation issue — it is a sovereignty issue. By developing and hosting a domestic large language model, Saudi Arabia reduces reliance on foreign AI infrastructure providers. This supports:

• National data residency objectives
• Regulatory alignment with PDPL
• Greater control over model development and updates
• Reduced exposure to geopolitical supply risks

The impact here is structural rather than transactional: HUMAIN represents a move from AI consumption to AI production within national borders.

3. Public-sector productivity potential
Although HUMAIN Chat was launched as a public-facing application, the features described in official materials — including real-time web search, bilingual switching between Arabic and English, dialect speech input, and conversation sharing — are clearly relevant to government workflows.

In practical terms, such a tool could support:

• Drafting correspondence and policy briefs in Arabic
• Translating between Arabic and English in ministries where bilingual documentation is required
• Citizen enquiry handling
• Knowledge retrieval and summarisation

Importantly, an independent evaluation of the underlying ALLaM-34B model provides early evidence of capability.

Researchers tested the system across a wide range of tasks: understanding Modern Standard Arabic, interpreting multiple regional dialects, switching naturally between Arabic and English within the same conversation, answering knowledge-based questions, solving reasoning problems, and maintaining safety standards.

In simple terms, the model was assessed not just on whether it could generate fluent sentences, but on whether it could:

• Understand different forms of Arabic accurately
• Follow complex instructions
• Switch languages smoothly when required
• Provide coherent and contextually appropriate answers

The evaluation found consistently strong performance across these areas, including particularly high scores in text generation and bilingual code-switching. For public institutions, that matters. Much of the administrative work depends on exactly these capabilities: drafting, summarising, translating and explaining.

What HUMAIN Chat represents is readiness. The infrastructure, the model, and the linguistic foundation are now in place. For public institutions, the question is no longer whether Arabic-native generative AI exists — but how it will be embedded into daily administrative practice.

Lessons learned

• Team Composition and Representation. Reporting notes a deliberate 50/50 gender balance within the team of 120 AI specialists. In a field where technical teams are often unevenly represented, this signals intentionality. Who builds an AI system influences how it interprets language, which edge cases are tested, and whose assumptions are challenged during development. Representation in design does not guarantee fairness in outcomes, but it shapes the foundation on which systems are built.
• Infrastructure and Data Governance. Infrastructure choices proved equally significant. HUMAIN is positioned as hosted entirely within national infrastructure and aligned with Saudi Arabia’s Personal Data Protection Law (PDPL). For governments, that is more than a technical configuration. Domestic hosting strengthens regulatory oversight, simplifies compliance, and reinforces public trust. Where data resides — and who governs it — becomes part of the policy story.
• Language Architecture and Model Design. The ALLAM model’s training depth, dialect coverage and evaluation across multiple forms of Arabic show that linguistic inclusion was engineered from the beginning. Cultural intelligence cannot be retrofitted onto a model trained primarily elsewhere. It must be built into the architecture — into the data, the testing, and the governance decisions.
• Implications for Institutional AI Development. In this instance, it was useful to embed language priorities, cultural context and legal frameworks at the design stage rather than layer them onto an existing foreign-built system.

This case study was written with assistance from artificial intelligence.

The application is fully compliant with the Saudi Personal Data Protection Law and is hosted entirely on HUMAIN's infrastructure within Saudi Arabia. Saudi Arabia's broader National AI Strategy and Vision 2030 framework provided the policy mandate and public investment rationale for the initiative. The Saudi Data and Artificial Intelligence Authority (SDAIA) played a role in developing the foundational model, giving the project an explicit regulatory and governmental lineage. Governments replicating this model should be aware that data sovereignty requirements — hosting on domestic infrastructure — significantly increase upfront capital costs but may be non-negotiable in politically sensitive public service contexts. The embedding of cultural and religious values at the model level, rather than through post-hoc content filtering, is a design decision with both technical and governance implications that requires cross-departmental input from legal, religious, and civil society stakeholders early in the process.