Falcon 2: UAE releases new AI model surpassing Meta’s Llama 3

UNITED ARAB EMIRATES: A government research institute in the United Arab Emirates (UAE) launched a new open-source Generative AI model on Monday, May 13.

Channel News Asia reported that Abu Dhabi’s Technology Innovation Institute (TII) introduced the Falcon 2 series.

This release includes two models: the text-based Falcon 2 11B and the vision-to-language Falcon 2 11B VLM, which can generate text descriptions from uploaded images.

The Technology Innovation Institute operates under the Advanced Technology Research Council of Abu Dhabi.

The UAE, known for its oil exports and regional influence, is now making significant investments in AI. However, the UAE’s foray into AI has not been without its challenges.

Last year, American officials issued a clear ultimatum to choose between American or Chinese technology. This prompted Emirati AI firm G42 to divest from Chinese companies and remove Chinese hardware from its operations.

The move paved the way for a US$1.5 billion (approx. S$2.03 billion) investment from Microsoft, facilitated in coordination with Washington.

Faisal Al Bannai, Secretary General of the Advanced Technology Research Council (ATRC) and president’s advisor on strategic research and advanced technology, stated that the UAE is proving it can be a significant player in artificial intelligence.

The Falcon 2 series was launched as companies and countries competed to create their own large language models following OpenAI’s release of ChatGPT in 2022.

While some have chosen to keep their AI code private, UAE’s Falcon and Meta’s Llama have made their code open-source for public use.

Al Bannai expressed optimism about Falcon 2’s performance and mentioned they are already working on the “Falcon 3 generation.”

Al Bannai says, “We’re very proud that we can still punch way above our weight, really compete with the best players globally.”

According to AFP, Falcon 2 11B has been tested against several leading AI models in its category of pre-trained models.

It surpasses the performance of Meta’s recently launched Llama 3 with 8 billion parameters and is nearly equal to Google’s Gemma 7B, achieving scores of 64.28 and 64.29, respectively.

Hugging Face, a US-based platform known for its objective evaluation tools and global leaderboard for open LLMs independently verified these performance results.

Falcon 2 11B is released under the TII Falcon License 2.0, a permissive licence based on Apache 2.0. This licence includes an acceptable use policy that encourages responsible AI use. To know more about Falcon 2, check here. /TISG

Read also: Apple and OpenAI to join forces to bring ChatGPT to iPhones!

Featured image by Depositphotos