For Saudi audit, law & professional firms
Which AI models handle Arabic documents best — and what can run on your own machines?
No single model wins on every Arabic document task. Frontier cloud models generally handle complex Arabic reasoning best, while Arabic-first open models such as ALLaM and SILMA can run locally on hardware your firm already owns and handle focused tasks — summarizing, extracting, drafting — well. For client work, the right choice depends on the task, your hardware, and, decisively, where the data is allowed to go.
Why Arabic Documents Trip Up Generic AI Tools
Most AI tools are trained overwhelmingly on English text, and it shows the moment you hand them a real Arabic contract or financial statement. Arabic is a root-and-pattern language: a single root can generate dozens of related words through prefixes, suffixes, and internal vowel changes, which makes tasks like search, extraction, and matching harder than in English. Diacritics that disambiguate meaning are often missing in real documents, right-to-left script has to coexist with embedded numbers, dates, and English terms, and everyday business writing mixes Modern Standard Arabic with dialect. None of this is exotic — it is normal for a Saudi firm's actual paperwork — but it is exactly the kind of detail that trips up a model that was mostly trained on English.
The practical effect: a tool that looks fluent in casual Arabic chat can still misread a clause, miscount a number, or garble formatting in a real financial statement or legal document. This is why firms that adopted a generic AI tool for English work often find it noticeably weaker once they point it at Arabic client documents.
The Model Landscape, in Plain Terms
Leaving aside marketing claims, the field breaks down into a few broad categories:
- Frontier cloud models (Claude, GPT-4-class models, and similar) — generally the strongest at complex Arabic reasoning and nuance, but they only run in the cloud, on the provider's infrastructure.
- Arabic-first open models such as ALLaM, developed in Saudi Arabia, and SILMA, from an Arabic-focused open-model team — built with Arabic morphology and usage in mind, and available in sizes small enough to run on hardware a firm already owns.
- Jais, an Arabic-English bilingual open model developed in the UAE, is another option built specifically for the language pair.
- Strong multilingual open families such as Qwen (Alibaba) and Llama (Meta) — trained across many languages including Arabic, with quality that varies significantly by model size and version.
None of these categories is universally best. Which one performs well depends heavily on the specific document, task, and language variety involved — which is why qualitative, hedged comparisons are more honest than a single ranking.
What Actually Matters When You're Choosing for a Firm
For a professional firm evaluating models for real client work, three questions matter more than any leaderboard:
- What is the task? Summarizing a contract, extracting figures from a financial statement, or drafting a routine letter is a different job than deep multi-step reasoning across a long, ambiguous document. Smaller local models tend to do well on the former and struggle more with the latter.
- What hardware do you actually have? Bigger models are usually more capable but need more computing power. A model that runs comfortably on a workstation with a strong GPU may be impractical on an ordinary office laptop.
- Where is the data allowed to go? This is the decisive factor for client work. A model can be technically excellent and still be the wrong choice if using it means sending a client's confidential contract or financial statement to a cloud provider's servers.
The Practical Pattern — and Honest Expectations for Local Models
In practice, the firms that get this right don't pick one model and use it for everything. They match the model to the prompt: sensitive client documents go to an Arabic-first model running on the firm's own machines, while non-sensitive, more complex work can be routed to a frontier cloud model when that produces a better result — decided automatically for each prompt rather than left to an employee's judgment in the moment.
It's worth being honest about what local models on ordinary office hardware can and can't do. Local Arabic-first models in common local sizes (roughly 7B-13B parameters) typically run more slowly than a cloud service, especially on CPUs rather than GPUs, and quantization — a common technique to make them fit on modest hardware — trades away some quality for speed and memory. What they do well is focused, bounded tasks: summarizing a document, extracting specific fields, drafting a first pass of routine text. What they do less well is long chains of complex reasoning across ambiguous, lengthy material. And unlike a cloud API, there is no per-use fee once the model is running on hardware you already own.
How to Evaluate Models on Your Own Documents
General comparisons are useful for narrowing the field, but Arabic quality is document-specific: a model's handling of formal legal Arabic, dialect-inflected client correspondence, and numeric financial tables can all differ. The most reliable way to choose is a short, structured pilot using your firm's own real documents — not a public benchmark, and not a demo built on someone else's sample contract.
A fixed-scope pilot lets a firm see, on its actual paperwork, which tasks a local Arabic-first model handles comfortably and which ones genuinely need a stronger cloud model — before committing to either approach across the whole office.
How Ithbat Applies This
Ithbat is built around this exact pattern rather than a bet on any single model. It installs on machines a firm already owns and runs Arabic-first and open models — including ALLaM and SILMA — locally through its own engine, alongside an OpenAI-compatible endpoint for one-click deployment across the firm's machines. Every prompt is classified before it is processed: prompts that touch Saudi sensitive data are kept and served on the firm's own hardware, on local Arabic-first models, and are never offered to a cloud provider. Non-sensitive prompts can route to a frontier cloud model such as Claude or ChatGPT, using the firm's own API key, with the routing decision logged either way. Each inference produces a tamper-evident, hash-chained Evidence Trace record, which feeds a bilingual Client Confidentiality Assurance report mapped to PDPL and NDMO. Engagements typically start with a fixed-scope, six-week pilot from SAR 15,000 for one office — enough time to test the pattern against a firm's own Arabic documents rather than a leaderboard.