Abu Dhabi’s Technology Innovation Institute (TII) announced three artificial-intelligence models for language, speech and document understanding on 6 October. It describes the 7-billion-parameter Falcon-Emirati as designed for Emirati Arabic vocabulary, idioms and cultural context. The second model, 1.6-billion-parameter Falcon-ASR, converts speech in Standard and Emirati Arabic, English, French, Spanish and Portuguese into text. Falcon-OCR-Arabic is designed to extract Arabic text and structured content from images and documents. Parameter count describes model size; by itself it does not establish accuracy, operating cost or commercial adoption.

TII says Falcon-Emirati scored 84.83% on Alyah, a benchmark for native Emirati Arabic, and outperformed the Arabic and multilingual open-source models it evaluated. The institute says Alyah covers everyday language, figurative expressions, heritage knowledge and poetry. This is a test result reported by TII, not an independent laboratory replication or proof of performance across every real-world use. TII says the model was built using native Emirati content, Modern Standard Arabic material about Emirati culture and heritage, and synthetic data informed by Emirati-specific linguistic resources. The announcement does not disclose dataset size, licences, error distribution or safety-evaluation results.

The speech model’s stated support for six languages creates possible uses in multilingual customer service, subtitles, meeting records and accessibility tools. TII also says it can mark when each word occurs in an audio file. The document model is intended to extract content including tables, mathematical formulas and document sections. These are described capabilities and potential applications, not reported productivity savings or customer deployments. Performance in practice will depend on audio quality, dialect variation, document image quality and the level of human review. A product announcement alone does not show whether accuracy is sufficient for high-impact decisions or regulated workflows.

Presenting the models together signals a product-family strategy built around regional language, speech and document needs. Specialisation may address gaps in general-purpose systems, while creating separate hosting, updating and evaluation work. For a company or public agency, procurement should therefore look beyond the announced benchmark score: error rates on its own tasks, total compute and integration costs, personal-data handling, training-data rights and the need for human checking all require assessment. TII says Falcon-Emirati will be available through Falcon Chat; the announcement does not describe the distribution terms for the other two models with the same level of detail.