IFM and Inception42 Unveil Jais 3, a Specialized Arabic AI Fleet

Built upon the K2 Horizon architecture released last month, Jais 3 features six language models ranging from 900 million to 375 billion parameters. Unlike general-purpose models that often struggle with regional shifts, this fleet is trained on Modern Standard Arabic and diverse dialects, including regional variations from the Gulf to North Africa. The system incorporates deep cultural context, ranging from traditional poetry to localized professional terminology, making it suitable for high-stakes environments like hospitals and government offices.

A significant component of the release is Jais Speech, a recognition engine that handles real-world conditions, including background noise and code-switching between Arabic, English, and French. During testing, the model outperformed existing commercial alternatives by maintaining lower error rates in transcriptions. The developers emphasize that Jais 3 is more than a translation tool; it is an agentic platform capable of executing multi-step workflows, such as researching data, drafting presentations, and managing complex spreadsheets.

Researchers at IFM and Inception42 evaluated the models against a custom suite of 24 benchmarks, finding that Jais 3 consistently leads open-source competitors in both linguistic fluency and instruction following. With a context window of 512,000 tokens—the largest for any Arabic-native model to date—the suite offers flexibility for developers to select sizes based on their specific computing hardware and operational budgets. The models, along with their training checkpoints and evaluation benchmarks, are currently available on Hugging Face.