ٰFrom UAE to World: Falcon LLM
Arabs are dedicated to resume civilisation.
It is no secret that Arabia is undergoing a major shift with focus on reclaiming the knowledge race that once was the theme of the region. One aspect of that is the efforts and attempts of research centres in United Arab Emirates and Kingdom of Saudi Arabia to conquer the AI field, which was charictarised in multiple acheivements such as Vicuna LLM, [mini-gpt4, a multi-modal LLM with visual capabilities]. Both are built on top of LLaMA by Meta. There's also Noor LLM from Technology Innovation Institute of UAE, which is a LLM model didcated for applications in Arabic language. While what is mentioned is so far good, the impressive and the surprising announcement was Falcon LLM from the same institute; An opne-source LLM using its own custom architecture and not based on LLaMa like most of open-source LLMs today. This is huge because legally Meta prevents users from using LLaMa or any model built on top of it commercially, while Falcon was announced with a royalty model; Companies making use of Falcon commercially and making an income beyond one million US dollars have to pay 10% fees to TII. This alone wasn't what made the name Falcon make rounds across the AI community, but also the results shown when released on HuggingFace allowing users to benchmark and test the model, down to the point where it has become clear Falcon has taken over LLaMa reign over open-source LLMs which it claimed for month, before HuggingFace officially giving the spot to Flacon in its Leaderboard. This is huge on different levels, one of which is the innovation and technical advancements are usually coming from faraway such as China or US, and seeing this breakthrough coming out to the world from Arabia is a sign of what is to come.
I'll leave technical details and overview to another post after I extensievly test the model, however I would end this post with this: Earlier I mentioned the model was launched with a royalty licensing scheme, and that was only until yesterday, as TII announced waiving the scheme and releasing the model freely under Apache license, which means the adoption of this model will take over beyond everything we saw so far.