TII launches Falcon-Emirati with new Arabic ASR and OCR models

The Technology Innovation Institute has launched three Falcon models focused on Emirati Arabic and multimodal language workflows. Falcon-Emirati is a 7-billion-parameter language model built around Emirati vocabulary, expressions, idioms and cultural context. It arrives alongside Falcon-ASR, a 1.6-billion-parameter speech-recognition model, and Falcon-OCR-Arabic, a compact model for extracting Arabic text and structured information from images and documents. The launch broadens the Falcon family from language generation into speech and document understanding.

TII says Falcon-Emirati was developed from Falcon-H1-Arabic using native Emirati content, cultural material and synthetic data informed by Emirati-specific linguistic resources. The institute reports a score of 84.83% on its Alyah native Emirati benchmark. Falcon-ASR supports Emirati Arabic, Modern Standard Arabic, English, French, Spanish and Portuguese and includes word-level timestamps. TII's Falcon model site describes Falcon-OCR-Arabic as a 270-million-parameter model. All comparative accuracy and benchmark claims remain attributed to TII rather than being treated as independent Open Model Weights measurements.

For the registry, this announcement requires artifact-level follow-up rather than automatic insertion. Falcon-Emirati, Falcon-ASR and Falcon-OCR-Arabic are separate models with different tasks and parameter scales, so their repositories, shard lists, formats, licenses and provenance must be verified individually. Until those exact primary artifacts are checked, the announcement is useful evidence of a publisher release but not a basis for filling unknown registry fields.