Alexandria-based AI company and nonprofit TokenAI has released Horus Taleeq 0.2B Base, an open-source Arabic-only language model for researchers and developers.
The model was trained from scratch on 6 billion Arabic tokens, with no other languages included. According to TokenAI, training took 30 hours on a single NVIDIA RTX PRO 6000 Blackwell Server Edition GPU.
TokenAI has also published the training code, data sources and supporting materials under the MIT License. Some training-data materials are still being uploaded.
Horus Taleeq is a pretraining base model rather than a finished chatbot. It has not undergone instruction tuning, preference optimisation or safety fine-tuning, so further work is required before production use. The company is positioning it as a foundation for continued pretraining, fine-tuning, dialect research and the development of instruction-tuned Arabic systems.
The release is TokenAI’s fourth major model launch this year and the first in a planned series focused on Arabic and its dialects. The company said future releases may include instruction-tuned checkpoints and broader evaluations.
Founded in 2025, TokenAI has also released Horus 1.0-4B, the multimodal Horus Hiero model for hieroglyphics and modern Arabic dialects, and Horus Cyber Nano.
Source: Middle East AI News
