Release of the Jais 3 models — open Arabic-focused agentic language models

A research team has introduced Jais 3, a series of six open Arabic-focused agentic language models ranging in size from 0.9B to 375B parameters. The models are not trained merely to answer text queries, but to complete multi-step tasks — planning, calling tools and MCP servers, working with the terminal, searching and reading the web, editing and creating PDF files, spreadsheets, and presentations, all while communicating in Arabic.

Rather than training from scratch, the authors adapted each model from the fully open K2 Horizon series for Arabic using continual pretraining, followed by supervised fine-tuning, final token preference optimization, and reinforcement learning with verified rewards (RLVR). The training data combines natural Arabic with translated, rewritten, and synthetic Arabic texts, supplemented with the base model's original training data to preserve its reasoning, agentic, and English-language capabilities. Jais 3 covers Modern Standard Arabic as well as major regional dialects and inherits the K2 Horizon model's 512K-token context window (across all models except the 0.9B version).

According to the authors, the 375B-parameter model is so far the largest open Arabic-focused language model. The team also reworked the methodology for evaluating Arabic models, switching to fully generative evaluation on audited Arabic benchmarks; according to the authors, Jais 3 achieves comparable or better results than open models of similar size across its size range on Arabic and agentic benchmarks, with stronger agentic capability, Arabic instruction following, and significantly fewer hallucinations.

The authors have released weights and checkpoints from all training stages for all six models, including samples of training data, data recipes, training configurations, training logs, audited benchmarks, and evaluation code.