Tarjama& deploys powerful Arabic model: Pronoia v2
New Arabic SLM aimed at enterprises outperforms other comparable Arabic models
#UAE #LLMs - Language technology and services provider Tarjama& has introduced Pronoia v2, an improved version of its Arabic-centric small language model (SLM). The new 14 billion parameter model significantly enhances Arabic language performance and expands its capabilities far beyond translation. Benchmarking conducted by Tarjama& shows Pronoia v2 outperforming GPT-4o, DeepSeek 70B, and Cohere 7B. The model scores high on a variety of evaluations, including the AlGhafa multiple-choice benchmark for Arabic LLMs. The model also offers high-speed processing on minimal hardware, making it a cost-effective and scalable on-premises AI solution for businesses and government organisations.
SO WHAT? - The release of Pronoia v2 strengthens Tarjama&’s position as an Arabic AI solution provider, addressing the gap in enterprise-level Arabic generative AI solutions. Many global AI models struggle with Arabic fluency and accuracy, wile Arabic-focused models often fall short when it comes to business use cases. The new version of its Pronoia SLM has been developed based on data and customer feedback gathered from pilot projects and implementations with enterprise customers. The improvements are therefore likely to prove to be valuable to customers.
Here are some more details about Pronoia v2:
Tarjama&, a UAE and Saudi Arabia-based language technology and services provider, has launched Pronoia v2, an enhanced version of its 14B parameter Arabic language model that was originally launched at GITEX 2024 last year.
Unlike larger AI models, Pronoia v2 runs on a single GPU, reducing costs and allowing for secure on-premises deployments.
Designed for enterprise-grade scalability, with support for Arabic and English, the model can handle long documents, chat interactions, and concurrent requests in stream mode.
Pronoia solutions incorporate advanced Retrieval-Augmented Generation (RAG) with verifiable citations for increased accuracy.
Features include an agentic tool for complex task execution and enterprise-level security for sensitive data.
Performance benchmark tests run by Tarjama& show Pronoia v2 outperforming GPT-4o, DeepSeek 70B, and Cohere 7B in Arabic tasks, including translation, summarisation, and rephrasing.
The model was tested on AlGhafa, a key Arabic LLM benchmark, and achieved top rankings in multiple-choice evaluations.
The Tarjama& development team also ran Pronoia v2 through the Open Arabic LLM Leaderboard benchmark evaluations, before removing it again. According to the team, it scored highly against the open source models in the leaderboard, ranking top with the highest average score.
Meanwhile, according to the company, customer evaluations have consistently rated Pronoia v2 higher than competitors in business-specific use cases, mainly where Arabic language text quality and accuracy matter.
ZOOM OUT - Arabic AI development has gained momentum in recent years, but many models fail to meet the business and enterprise-grade quality demands of the MENA (Middle East and North Africa) region. Tarjama&’s high-quality Arabic datasets and its focus on business, legal, and medical AI applications differentiate Pronoia v2 from general-purpose LLMs. As regional demand for AI-driven automation, translation, and enterprise intelligence rises, Tarjama& could play a key role in setting expectations for what enterprise Arabic GenAI should look like.
[Written and edited with the assistance of AI]
Read more about Arabic large language models:
Mistral AI unveils Mistral Saba for Arab world & South Asia (Middle East AI News)
MBZUAI releases Arabic 'inclusive' multimodal AI model (Middle East AI News)
Qatar launches Fanar sovereign large language model (Middle East AI News)
Tarjama& launches LLM for business Arabic translation (Middle East AI News)