Falcon (TII)
Abu Dhabi TII's open Falcon models: hybrid Transformer-Mamba Falcon-H1 (0.5B to 34B, 256K context), Falcon-H1R 7B reasoning, and small OCR and vision models
Overview
Falcon is the open model family of the Technology Innovation Institute (TII), the applied research arm of Abu Dhabi's Advanced Technology Research Council, founded in 2020. Its current models use a hybrid design that combines Transformer attention with Mamba-2 state-space layers, an approach aimed at efficient long-context work. The Falcon-H1 family spans 0.5B, 1.5B, 1.5B-Deep, 3B, 7B and 34B models with up to 256K context, and Falcon-H1R 7B, released in January 2026, is a reasoning model that TII says matches models several times its size. Newer uploads on Hugging Face are small special-purpose models such as Falcon-OCR (0.3B) and Falcon-Perception (0.6B and 300M), plus 1.58-bit Falcon3 and Falcon-E models built for CPUs. The family also includes Falcon Mamba 7B, Falcon 3, Falcon Arabic and Falcon-Emirati 7B. Licenses vary: the Falcon-H1 GitHub repo carries an Apache 2.0 badge, while the Falcon-H1R model card uses the Falcon-LLM License with an acceptable use policy TII can update. TII lists no hosted API on its site, so you run Falcon yourself from the weights.
The Verdict
Who Should Use Falcon?
Best For
- Small, efficient reasoning on one GPU (Falcon-H1R 7B)
- Long-context work on modest hardware (hybrid Mamba design)
- Arabic and Emirati dialect applications
- CPU and edge deployment (Falcon-E, 1.58-bit models)
Not Ideal For
- Teams that want a managed API from the model maker
- Frontier-scale tasks (largest current H1 model is 34B)
- Buyers who want a large community and many third-party reviews
What's Great
- Hybrid Transformer and Mamba-2 design built for long context
- Falcon-H1R 7B scored 34% on Terminal-Bench Hard in one independent test, above DeepSeek R1 Qwen3 8B (26.9%)
- TII claims up to 1,500 tokens a second per GPU for H1R
- Wide size range, down to sub-1B OCR and vision models
Watch Out For
- "Beats models 7x its size" claims rest on selective, vendor-run benchmarks
- A 2026 review found the hybrid overhead adds no capability gain for math and science
- Custom Falcon-LLM License on H1R; use policy can change
- Modest adoption: H1R had about 2,600 Hugging Face downloads last month
Pricing
View all features & details
Current Models
- Falcon-H1R 7B: hybrid reasoning model (Jan 2026)
- Falcon-H1: 0.5B, 1.5B, 1.5B-Deep, 3B, 7B, 34B, up to 256K context
- Falcon-OCR (0.3B) and Falcon-Perception (0.6B, 300M)
- Falcon-E and Falcon3 1.58-bit models for CPUs
- Falcon Arabic, Falcon-Emirati 7B, Falcon Mamba 7B
Older Models
- Falcon 2 (11B, vision-to-language)
- Falcon 40B under Apache 2.0
- Falcon 180B under a royalty-free license based on Apache 2.0 with restrictions
Organization
- Technology Innovation Institute, Abu Dhabi
- Applied research pillar of the Advanced Technology Research Council (ATRC)
- Founded 2020, unveiled November 2020
Real-World Usage
Independent Tests
- Towards AI: H1R 7B at 34% on Terminal-Bench Hard, just above Qwen3-32B (33.4%)
- Layer3 Labs: solid for size, but the hybrid design gives no clear gain on math and science
Activity
- TII keeps publishing to Hugging Face, including Falcon-OCR in late 2026
- The Falcon-H1 GitHub repo was last pushed in October 2025
How It Compares
| Feature | Falcon-H1R 7B | Phi-4-mini-flash-reasoning | Ministral 3 8B | Jamba Reasoning 3B |
|---|---|---|---|---|
| Maker | TII (Abu Dhabi) | Microsoft | Mistral AI | AI21 Labs |
| Parameters | 7B | 3.8B | 8B | 3B |
| License | Falcon-LLM License | MIT | Apache 2.0 | Apache 2.0 |
| Context | Up to 256K (H1 family) | 64K | Not stated | 256K |
| Hosted API by maker | None listed | Azure AI Foundry (Phi family) | Yes, $0.15 per 1M | Not stated |
Sources: Falcon LLM · Hugging Face · Falcon-H1R 7B · Falcon-H1 GitHub · The Decoder
Is this your tool? Add a badge