Technology Innovation Institute

A government-backed research centre

The Technology Innovation Institute (TII) describes itself as a global research centre and as part of the Abu Dhabi Government's Advanced Technology Research Council^[1]. It does not concentrate on a single discipline: it lists ten focus areas, among them artificial intelligence and digital science, autonomous robotics, cryptography, directed energy, propulsion and space, quantum, and renewable and sustainable energy^[1].

That breadth shapes how the institute presents its work. TII says it runs both partnership research with international institutes and proprietary research across its focus areas^[1], so cryptography work and language-model training sit inside the same organisation rather than in separate companies.

Leadership

TII lists Dr. Najwa Aaraj as chief executive officer and Dr. Chaouki Kasmi as chief innovation officer^[1]. Its artificial intelligence work is organised into a named centre: Dr. Hakim Hacid is listed as chief researcher of the Artificial Intelligence and Digital Science Research Center^[1], one of the focus areas turned into an organisational unit with a research lead of its own.

The Falcon models

Outside the institute, TII is best known for the Falcon family of language models. On 14 December 2024 it published the Falcon3 models in four sizes — 1B, 3B, 7B and 10B — each in a base form and an instruct form^[2]. The two forms do different jobs: a base model is trained to continue text, while an instruct model is tuned afterwards to follow a user's request, which is why every size appears twice in the release.

The 10B model was not designed and trained from scratch at that size. It was produced by depth up-scaling a 7B base model, trained over 2 trillion tokens on 1,024 H100 GPUs^[2]. Depth up-scaling means growing a smaller model into a larger one by extending its depth and continuing training, instead of starting again with a new architecture of the target size.

On 1 May 2025 TII released the weights of Falcon-H1^[3]. It was the first Falcon family to run Transformer attention in parallel with state space models^[3] — two different mechanisms for handling a sequence, active side by side inside one network. Attention compares each position with the others, which is strong at picking out precise details but expensive on long inputs; a state space model compresses the sequence as it moves along, which is cheaper. Running the two in parallel is an attempt to keep both properties at once. Falcon-H1 shipped in six sizes from 0.5B to 34B, and its technical report followed on 30 July 2025^[3].

Licensing

The terms of release are TII's own. The Falcon-LLM Licence is a custom licence published on TII's own site rather than an OSI licence^[2][4]. An OSI licence is one approved by the Open Source Initiative, the body that maintains the standard list of open-source licences. Model weights are not ordinary software, and licences written for them are frequently drafted by the organisation doing the release; whether a given release counts as open source therefore depends on what its terms actually permit.

Where it sits

TII belongs to a group of government-funded laboratories that publish model weights and technical reports rather than keeping models in-house: the Falcon3 models were published for others to use^[2], the Falcon-H1 weights were released^[3], and a technical report describing the latter followed roughly three months later^[3]. The releases also show a pattern in how the models are built — a smaller base grown into a larger one for Falcon3, and two sequence mechanisms combined for Falcon-H1 — with the licence terms set by the institute itself^[2][4].