ALLaM

ALLaM (Arabic Large Language Model) is the Saudi Data and AI Authority's sovereign LLM program, developed by the National Center for AI. The family has two tracks: models pretrained from scratch (notably a 7B trained on 4T English tokens followed by 1.2T mixed Arabic/English tokens, transferring English capability into Arabic without catastrophic forgetting) and adapted models continuing from open weights (a 13B built on Llama-2 with expanded Arabic vocabulary). SDAIA launched ALLaM as open-source (with optional acceptable-use policy) on IBM watsonx in May 2024.

Limitations: with a 4K context window and 2024-era scale, ALLaM is not competitive with 2026 frontier models on general tasks — its value is Arabic specialization and sovereignty, not raw capability. The licensing (open-source with optional AUP) also needs case-by-case review for commercial deployment. For Arabic-first applications, government work in the Gulf, and research into cross-lingual transfer, it remains the flagship sovereign Arabic model.