hammh0a/Hala-1.2B

TEXT GENERATIONPricing:Input $0.04 / Cached $0.008 / Output $0.08Concurrent Unit Cost:1Model Size:1.2BQuant:BF16Context Size:32kPublished:Sep 13, 2025License:cc-by-nc-4.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

Hala-1.2B is a 1.2 billion parameter, Arabic-centric instruction and translation model developed by Hasan Abed Al Kader Hammoud, Mohammad Zbeeb, and Bernard Ghanem from King Abdullah University of Science and Technology (KAUST). This model is specifically optimized for Arabic language tasks, demonstrating strong performance across various Arabic benchmarks. With a context length of 32768 tokens, Hala-1.2B is designed for applications requiring high-quality Arabic language understanding and generation, outperforming other models in its size category on average Arabic benchmarks.

Loading preview...

Hala: Arabic-Centric Instruction & Translation Models

Hala-1.2B is a 1.2 billion parameter model developed by researchers at King Abdullah University of Science and Technology (KAUST), including Hasan Abed Al Kader Hammoud, Mohammad Zbeeb, and Bernard Ghanem. This model is part of the Hala family, which focuses on building high-performing instruction and translation models specifically for the Arabic language. The name "Hala" (حلا) signifies sweetness and beauty in Arabic, reflecting the project's dedication to the language.

Key Capabilities

  • Superior Arabic Language Performance: Hala-1.2B demonstrates leading performance in its size category (≤2B parameters) across a suite of Arabic benchmarks, achieving an average score of 51.4. This includes strong results on AlGhafa (59.2), ArabicMMLU (48.6), EXAMS (43.4), MadinahQA (41.6), AraTrust (71.7), and ArbMMLU-HT (44.2).
  • Instruction Following: The model is designed to follow instructions effectively, making it suitable for various NLP tasks.
  • Translation: Optimized for translation tasks, leveraging its Arabic-centric training.
  • Extended Context Length: Features a substantial context length of 32768 tokens, allowing for processing longer texts and more complex queries.

Good For

  • Arabic NLP Applications: Ideal for developers and researchers working on applications that require robust Arabic language understanding, generation, and translation.
  • Instruction-tuned tasks in Arabic: Excels in scenarios where precise instruction following in Arabic is critical.
  • Benchmarking Arabic LLMs: Serves as a strong baseline and competitive model for evaluating other Arabic language models, particularly in the ≤2B parameter range.
  • Research in Arabic AI: A valuable resource for academic and industrial research focused on advancing AI capabilities for the Arabic language.

Finetunes
2 models