The UAE releases Falcon 40B under an open licence

The Technology Innovation Institute (TII), a research body funded by the government of Abu Dhabi, released Falcon-40B, a 40-billion-parameter language model, under the Apache 2.0 licence — permitting commercial use with no royalties or restrictions. Training had finished in February 2023 after two months on a custom dataset called RefinedWeb, a filtered and deduplicated web corpus supplemented with curated sources, totalling roughly 1 trillion tokens.

TII’s own model card claimed Falcon-40B outperformed LLaMA, StableLM, RedPajama and MPT, describing it as “the best open-source model currently available” and pointing to Hugging Face’s Open LLM Leaderboard for verification — where it took the top spot on release. Like other base models of the period, TII noted it required further fine-tuning for most practical uses, and released an instruct variant alongside it.

The release mattered less for the architecture, which followed the by-then-standard GPT-style decoder recipe, than for who had built it. Falcon came from a state-funded institute in the Gulf rather than a Silicon Valley lab, a Chinese tech company, or an EleutherAI-style volunteer collective — evidence that competitive open-weight models no longer required either venture funding or a national AI champion, and that the permissive-licence, openly-benchmarked release pattern set by LLaMA and RedPajama earlier that year was being adopted well beyond the labs that started it.