Most are built for English and struggle with Arabic, especially handling tasks beyond simple question-answering. Arabic speakers need models that can plan multi-step tasks, use tools, and produce documents in their own language and dialects.
The researchers adapted six existing open models (from 0.9 billion to 375 billion parameters) to Arabic by continuing their training on Arabic text, natural, translated, and synthetic, while preserving their original reasoning and English skills. They fine-tuned the models on instructions and reinforced them with verifiable rewards, then evaluated them using audited Arabic benchmarks.
The largest model is the biggest open Arabic-centric model to date. Across all sizes, Jais 3 matches or outperforms comparable open models on Arabic and agentic tasks, with particularly strong instruction following and far fewer incorrect statements. The team released all model weights, training data, and evaluation code.