النماذج متعددة اللغات والعربية: ماذا تختار للمؤسسة؟

Landscape 2025–2026

Multilingual frontier models

Provider	Model	Arabic notes
OpenAI	GPT-4o / o-series	Strong MSA; API
Anthropic	Claude 3.5+	Long context, good AR
Google	Gemini 1.5+	Native multilingual
Meta	Llama 3.x	Open, self-host
Alibaba	Qwen 2.5	Strong multilingual OSS

Arabic-focused / regional

Model	Origin	Focus
Jais	UAE	Arabic + English, open weights tiers
ALLaM	KSA (SDAIA)	Arabic instruction
AceGPT	research	Arabic chat
Fanar	Qatar	Islamic/Arabic context

ACL / ArabicNLP workshops publish benchmarks — FLARE, ALMEval.

Decision matrix

Factor	Multilingual API	Arabic regional	Self-host Llama/Qwen
Data residency	محدود	نعم	نعم
AR dialect	medium	high	medium-high
EN quality	high	medium	high
Cost at scale	OpEx	varies	CapEx GPU
Time to pilot	days	weeks	weeks

Benchmark methodology

Don't trust single BLEU score. Test your tasks:

1. Summarize 20 Arabic policy PDFs — factual accuracy
2. Extract JSON fields from Arabic invoices
3. QA 50 employee questions (MSA + dialect mix)
4. Translate AR↔EN paragraph — domain terms

Score: human rubric 1–5 + source grounding if RAG.

Token economics

Arabic script → more UTF-8 bytes → higher token count vs English same meaning.

Example (approximate):

• English 100 words ≈ 130 tokens
• Arabic 100 words ≈ 200–250 tokens

Budget impact: 1.5–2x cost vs English-only org at same volume.

Mitigations:

• concise prompts in Arabic
• RAG (shorter context)
• smaller model for routing

Prompting multilingual systems

Language: respond in Modern Standard Arabic unless user writes dialect;
match user dialect only for clarification, then answer in MSA.
Technical terms: provide Arabic with English in parentheses on first use.

Hybrid strategy (recommended MENA enterprise)

Router (small model) → detect lang/intent
        ↓
   AR query → RAG (Arabic index) → multilingual LLM → AR answer
   EN query → RAG (EN index or same with lang filter)

Separate indexes if translation quality insufficient.

Sovereignty note

Gov/ministry: prioritize Jais/ALLaM/Llama on sovereign cloud
Private enterprise: GPT/Claude acceptable for non-sensitive with DPA

المصادر

• Sengupta et al., Jais paper (2023)
• SDAIA ALLaM announcements
• ArabicNLP / WANLP shared tasks
• OpenAI — multilingual performance documentation