The Technology Innovation Institute in the United Arab Emirates announced today, Monday, the launch of the second version of its large language model named "Falcon 2," which outperforms Meta's new Llama 3 model, according to a statement on its website. The institute, recognized as a leading global center for scientific research and a key pillar of applied research under the Advanced Technology Research Council in Abu Dhabi, stated that "the new release includes two advanced versions: Falcon 2 11B, a more efficient large language model trained on 5.5 trillion tokens with 11 billion parameters, and the second version, Falcon 2 11B (VLM), which features 'vision-to-language' capabilities that allow for effective conversion of visual inputs into textual outputs."
It is noteworthy that both models are equipped with multilingual functionality, and the Falcon 2 11B (VLM) is the first multimedia model launched by the Technology Innovation Institute. Additionally, it is the only model in the market with the ability to convert visual images into written text, representing a significant advancement in AI innovations.
The Technology Innovation Institute, part of the Advanced Technology Research Council in Abu Dhabi, continues to enhance its efforts to build a sustainable operational framework for generative AI projects, having launched the Falcon Foundation focused on the development of open-source generative AI models.
Falcon 2 11B has been tested against several prominent AI models within the same category, showing superior performance over Meta's new Llama 3 by 8 billion parameters (8B), and similar performance to Google's Gemini 7B (Falcon 2 11B: 64.28 vs. Gemini 7B: 64.29), according to a list published by Hugging Face, a leading platform in the U.S. that utilizes objective evaluation tools to assess open-source large language models.
Importantly, both Falcon 2 11B and 11B (VLM) are open-source, allowing developers worldwide to use them with unlimited access. Future expansions and developments of the next generation of Falcon 2 models are planned to offer a variety of sizes and ranges. These models will be enhanced with advanced machine learning capabilities like the "Mixture of Experts" (MoE) system, aiming to elevate performance to more advanced levels.
All AI models released by the Technology Innovation Institute so far have been ranked among the highest globally and are considered some of the best open-source language models. The newly developed Falcon 2 11B models are expected to better meet market needs, particularly in the continuously evolving domain of generative AI.
Falcon 2 11B models feature multilingual capabilities, effectively handling tasks in English, French, Spanish, German, Portuguese, and many other languages, enriching their diversity and increasing their effectiveness in various scenarios. The Falcon 2 11B (VLM) includes a "vision-to-language" system capable of identifying and analyzing visual images, offering a wide range of applications across multiple sectors such as healthcare, finance, e-commerce, education, and legal sectors, thereby supporting document management, digital archiving, visually impaired individuals, and other valuable uses. Moreover, these models can be efficiently run on a single graphics processing unit (GPU), making them scalable and easy to deploy and integrate into lightweight infrastructures like laptops and other small devices.
