Which Falcon Models Are on the TII Site and What Is Each One For?

Updated

The Falcon site (falconllm.tii.ae) presents a family of generative AI models rather than a single product. Each model targets a different job: general language and reasoning, Arabic language, image understanding, Emirati dialect, speech transcription, and Arabic text extraction from images. The table below maps each model to its stated purpose so you can pick the right starting point.

Model-by-model overview

Model What the site says it does Best fit when you need…
Falcon-H1 Hybrid architecture combining Transformer and Mamba designs; positioned as removing the trade-off between performance and efficiency General-purpose generation where you care about both output quality and compute cost
Falcon H1R 7B Compact 7-billion-parameter model with advanced reasoning in mathematics and coding; combines attention-based and state-space architectures for efficient sequence modeling under large batch sizes and long responses Reasoning-heavy tasks (math, code) at a smaller model size, or serving many long requests at once
Falcon-H1-Arabic High-performance Arabic model described as a departure from traditional LLMs Arabic-language applications that need strong performance, not just translation
Falcon Arabic Built for Modern Standard Arabic and dialects Work spanning formal Arabic and regional dialects
Falcon Perception Multimodal model that lets systems see, read, and understand images through natural-language prompts Image understanding driven by plain-language instructions rather than fixed labels
Falcon-Emirati 7-billion-parameter model purpose-built for Emirati vocabulary, expressions, and cultural context Content or services aimed at Emirati users where local phrasing matters
Falcon-ASR Emirati and Arabic speech transcription, supporting multilingual voice services Converting spoken Arabic or Emirati speech into text
Falcon-OCR-Arabic Lightweight model for extracting text from images, with leading Arabic extraction performance Pulling Arabic text out of scans, photos, or documents

How the pieces relate

The site frames these as one family under the banner of responsible, accessible AI. Two groupings help:

  • Language and reasoning: Falcon-H1 is the general foundation; H1R 7B is the compact reasoning variant; Falcon-H1-Arabic and Falcon Arabic cover Arabic, with Falcon-Emirati narrowing further to Emirati usage.
  • Perception and input: Falcon Perception handles images generally, while Falcon-ASR (speech) and Falcon-OCR-Arabic (image text) are specialized input models for Arabic and Emirati contexts.

Choosing between them

  • If your task is text generation in English or general languages, start with Falcon-H1; consider H1R 7B if math/coding reasoning or high-volume long responses matter.
  • If your content is Arabic, choose Falcon Arabic for dialect coverage or Falcon-H1-Arabic for performance-focused Arabic work; use Falcon-Emirati when local UAE phrasing is the priority.
  • If your input is not text, use Falcon Perception for general image understanding, Falcon-ASR for speech, and Falcon-OCR-Arabic for extracting Arabic text from images.

The site describes purposes and architectures but does not publish pricing or access terms on the pages summarized here, so confirm availability and licensing on each model's page before committing.