RightNow-Arabic-0.5B-Turbo is a new 518M-parameter Arabic-specialized decoder LLM, built on Qwen2.5-0.5B, designed to bridge the gap between small multilingual and large Arabic-specialized models. Its development pipeline included adding 27,032 Arabic tokens via vocabulary injection, continued pretraining on 504M Arabic tokens, and fine-tuning with supervised instruction and direct preference optimization. The model achieved a 35.9% mean accuracy on three Arabic benchmarks (COPA-ar, Arabic HellaSwag, ArabicMMLU), outperforming all same-class open models and recovering 67% of SILMA-9B's mean accuracy at 1/18 the parameters, with all code and weights publicly released. Why it matters: This model significantly advances efficient Arabic NLP by providing a powerful, specialized sub-1B LLM suitable for edge deployment, making advanced Arabic AI more accessible and performant on resource-constrained devices.
Researchers at MBZUAI have introduced EvoLMM, a self-evolving framework for large multimodal models that enhances reasoning capabilities without human-annotated data or reward distillation. EvoLMM uses two cooperative agents, a Proposer and a Solver, which generate image-grounded questions and solve them through internal consistency, using a continuous self-rewarding process. Evaluations using Qwen2.5-VL as the base model showed performance gains of up to 3% on multimodal math-reasoning benchmarks like ChartQA, MathVista, and MathVision using only raw training images.
The authors introduce Nile-Chat, a collection of LLMs (4B, 3x4B-A6B, and 12B) specifically for the Egyptian dialect, capable of understanding and generating text in both Arabic and Latin scripts. A novel language adaptation approach using the Branch-Train-MiX strategy is used to merge script-specialized experts into a single MoE model. Nile-Chat models outperform multilingual and Arabic LLMs like LLaMa, Jais, and ALLaM on newly introduced Egyptian benchmarks, with the 12B model achieving a 14.4% performance gain over Qwen2.5-14B-Instruct on Latin-script benchmarks; all resources are publicly available. Why it matters: This work addresses the overlooked aspect of adapting LLMs to dual-script languages, providing a methodology for creating more inclusive and representative language models in the Arabic-speaking world.
This paper introduces two methods for creating Arabic LLM prompts at scale: translating existing English prompt datasets and creating natural language prompts from Arabic NLP datasets. Using these methods, the authors generated over 67.4 million Arabic prompts covering tasks like summarization and question answering. Fine-tuning a 7B Qwen2 model on these prompts outperforms a 70B Llama3 model in handling Arabic prompts. Why it matters: The research provides a cost-effective approach to scaling Arabic LLM training data, potentially improving the performance of smaller, more accessible models for Arabic NLP.
MBZUAI's Institute of Foundation Models has released K2, a 70-billion-parameter, reasoning-centric foundation model. K2 is designed to be fully inspectable, with open weights, training code, data composition, mid-training checkpoints, and evaluation harnesses. K2 outperforms Qwen2.5-72B and approaches the performance of Qwen3-235B. Why it matters: This release promotes transparency and reproducibility in AI development, providing researchers with the resources needed to study, adapt, and build upon a strong foundation model.
A new paper from MBZUAI demonstrates that state-of-the-art speech models can be easily jailbroken using audio perturbations to generate harmful content, achieving success rates of 76-93% on models like Qwen2-Audio and LLaMA-Omni. The researchers adapted projected gradient descent (PGD) to the audio domain to optimize waveforms that push the model towards harmful responses. They propose a defense mechanism based on post-hoc activation patching that hardens models at inference time without retraining. Why it matters: This research highlights a critical vulnerability in speech-based LLMs and offers a practical solution, contributing to the development of more secure and trustworthy AI systems in the region and globally.
Researchers fine-tuned the Qwen2-1.5B model for Arabic using QLoRA on a 4GB VRAM system, using datasets like Bactrian and Arabic Wikipedia. They addressed challenges in Arabic NLP including morphology and dialectal variations. After 10,000 training steps, the final loss converged to 0.1083 with improved handling of Arabic-specific linguistic phenomena. Why it matters: This demonstrates a resource-efficient approach for creating specialized Arabic language models, democratizing access to advanced NLP technologies.