Editing Arabic files and using software tools are central tasks for Jais 3, a new family of six open models. MBZUAI's Institute of Foundation Models released the family on October 15, with sizes from 900 million to 375 billion parameters.
Arabic support often means producing fluent replies. This release also targets agents, programs that use a model to plan several steps and call tools. The developers describe workflows involving documents, spreadsheets, web searches and terminal commands.
Jais 3 supports Modern Standard Arabic, regional dialects and Arabizi, Arabic written in Latin letters. Paired with the separate Jais-ASR speech-recognition model, it can accept spoken instructions before attempting a task.
Most sizes have a context window of 512K tokens, the amount of material available to a model during a run. The smallest has 128K. A large window does not itself establish accurate reasoning over a long document.
The institute reports strong Arabic and tool-use results in its own evaluations. It also says it audited Arabic benchmarks and changed scoring procedures. Those details matter when comparing the published numbers with results produced under other test setups.
The largest version uses 23 billion active parameters out of 375 billion total. Smaller versions target more modest hardware, although the actual requirements depend on the model and numerical format.
Weights, training-stage checkpoints and evaluation code are available under Apache 2.0. The release makes the models inspectable and reusable. It does not certify every tool action they take.
