TII

nonprofit

uae · Founded 2020

The Technology Innovation Institute (TII) is the applied research pillar of the Abu Dhabi government's Advanced Technology Research Council (ATRC), established in 2020. Led by CEO Dr. Najwa Aaraj (PhD Princeton) with AI research directed by Prof. Merouane Debbah (96K+ citations, H-index 98) and Dr. Hakim Hacid. Over 45M total Falcon downloads.

The Falcon series evolved from dense Transformers (Falcon 7B/40B/180B, 2023, trained on RefinedWeb) through Falcon 2 (11B, 5.5T tokens) and Falcon 3 (Dec 2024, up to 10B, 14T tokens) to the current hybrid architecture: Falcon-H1 (May 2025, parallel Transformer + Mamba-2 heads, 0.5B–34B, 256K context) and Falcon-H1R (Jan 2026, 7B reasoning, AIME25 83.1%, AA Intelligence 16). H1-34B matches 70B-class models with up to 8x output throughput improvement.

TII also produced the first competitive pure-Mamba LLM (Falcon Mamba 7B, 5.8T tokens, constant-memory inference at 130K+) and Falcon Perception (multimodal vision, 600M params matching Meta SAM3). All Falcon models are open-weight.

Website Wikipedia HuggingFace GitHub

open-weightfrontierarchitecturemultilingual

People

Merouane Debbah Google Scholar OpenReview — Chief Researcher, AI Cross-Center Unit (formerly Professor, CentraleSupélec / Khalifa University)
Hakim Hacid Google Scholar OpenReview — Chief Researcher, AI and Digital Science Research Center (formerly Bell Labs; Zayed University)
Ebtesam Almazrouei — Former Acting Chief AI Researcher (now CEO of AIE3)

Outputs (7)

Falcon-H1R

model

2026-01-05

7B hybrid reasoning model (Transformer + Mamba-2 parallel heads). Trained via SFT on 3.1M samples + GRPO RL. 256K context, up to 48K response token…

Falcon-H1

model

2025-05-21

Hybrid-head architecture: Transformer attention and Mamba-2 SSM heads run in parallel (concatenated), not interleaved. 0.5B to 34B…

Falcon 3

model

2024-12-01

Family of 1B, 3B, 7B, 10B dense Transformers plus Mamba-7B SSM variant. 7B trained from scratch on 14T tokens (1,024 H100s). 10B created via depth ups…

Falcon Mamba

model

2024-10-07

First competitive pure-Mamba (attention-free) 7B LLM. 64 layers, 4096 hidden dim, trained on 5.8T tokens using 256 H100s. Progressive context (2K→8K t…

Falcon 2

model

2024-07-20

11B dense Transformer trained on 5.5T tokens in 4 stages with progressive context extension (2K→8K). 11 languages. VLM variant adds CLIP ViT-L/14 visi…

Falcon (7B / 40B / 180B)

model

2023-11-28

Original Falcon series. Dense causal Transformers with multi-query attention, RoPE, FlashAttention. Trained primarily on <a href="https://arxiv.org…

RefinedWeb

dataset

2023-06-01

~5 trillion token open web corpus from CommonCrawl demonstrating that properly filtered web data alone can outperform curated corpora. 600B token publ…