The UAE has two major open-source AI model families: Falcon (from TII) and Jais (from G42/MBZUAI). Both are open source, both are well-funded, but they serve very different purposes.
Head-to-head
| Falcon | Jais | |
|---|---|---|
| Developer | Technology Innovation Institute (TII) | G42 / MBZUAI / Cerebras |
| Focus | General purpose, multilingual | Arabic-first, bilingual |
| Flagship | Falcon 2 11B / H1R 7B | Jais 2 70B |
| Smallest | Falcon H1R 7B | Jais 2 8B |
| Arabic quality | Good | ✅ Best |
| English quality | ✅ Strong | Good |
| Coding | Good (H1R for reasoning) | Basic |
| Reasoning | ✅ Strong (H1R hybrid) | Good |
| Architecture | Hybrid SSM + attention (H1R) | GPT-3 decoder |
| License | Apache 2.0 | Open weights |
| Run locally | ✅ Ollama | ✅ Ollama / HuggingFace |
When to use Falcon
- English-first applications with multilingual support
- Reasoning-heavy tasks (math, logic, coding) — Falcon H1R-7B
- Budget hardware — H1R-7B runs on 6GB RAM
- General-purpose AI — chatbots, content generation, analysis
When to use Jais
- Arabic-first applications — customer support, content, documentation
- Arabic dialect support — Egyptian, Gulf, Levantine, Moroccan
- Arabic-English code-switching — mixed language conversations
- Cultural context — understands Arabic cultural references and idioms
For developers specifically
| Task | Best pick | Why |
|---|---|---|
| Coding in English | Falcon H1R 7B | Better reasoning |
| Arabic documentation | Jais 2 | Native Arabic quality |
| Arabic chatbot | Jais 2 | Dialect support |
| Multilingual app | Falcon 2 11B | Broader language coverage |
| Math/reasoning | Falcon H1R 7B | Hybrid architecture |
Both vs Chinese models
For coding specifically, both UAE models fall behind the Chinese open-source leaders:
| Model | Coding | Reasoning | Arabic |
|---|---|---|---|
| Qwen3 8B | ✅ Better | Good | Decent |
| Yi-Coder 9B | ✅ Best | Decent | Decent |
| DeepSeek R1 14B | Good | ✅ Best | Decent |
| Falcon H1R 7B | Good | Strong | Good |
| Jais 2 8B | Basic | Good | ✅ Best |
If coding is your primary use case, the Chinese models (Yi-Coder, Qwen, DeepSeek) are stronger. Falcon and Jais shine in their specific niches: reasoning (Falcon H1R) and Arabic (Jais).
The technical differences that matter
Tokenizer efficiency
This is the biggest practical difference. Jais uses a custom Arabic-centric vocabulary designed from scratch. General-purpose tokenizers (used by Falcon, Llama, GPT) were built for English and treat Arabic characters inefficiently — often splitting a single Arabic word into 3-5 tokens.
Jais’s tokenizer handles Arabic at roughly 2-3x the efficiency. In practice, this means:
- The same 8K context window holds 2-3x more Arabic text in Jais vs Falcon
- Arabic inference is faster (fewer tokens to process)
- Arabic generation is more coherent (the model “thinks” in Arabic-native tokens)
For English text, both tokenizers perform similarly.
Architecture
Falcon H1R 7B uses a hybrid Transformer-Mamba (State Space Model) architecture. This gives it:
- 256K context window (vs 8K for standard Falcon 2)
- 1,500 tokens/second per GPU at batch size 64
- 2x the throughput of similarly-sized transformers like Qwen3-8B
- Linear scaling with sequence length
Jais 2 uses a standard GPT-3 decoder-only architecture. It’s not architecturally innovative — its strength is the training data (126B Arabic tokens, 251B English, 50B code) and the Arabic-optimized tokenizer.
Dialect support
| Dialect | Jais 2 | Falcon 2 |
|---|---|---|
| Modern Standard Arabic (MSA) | ✅ Excellent | Good |
| Egyptian Arabic | ✅ Good | Basic |
| Gulf Arabic | ✅ Good | Basic |
| Levantine Arabic | ✅ Good | Basic |
| Moroccan Arabic | ✅ Decent | Poor |
| Code-switching (Arabic + English) | ✅ Good | Decent |
Jais was specifically trained on dialectal Arabic data. Falcon was trained on 11 languages with Arabic as one of many, so dialect support is limited.
Deployment considerations
| Factor | Falcon | Jais |
|---|---|---|
| Ollama support | ✅ Official (falcon2) | Community models |
| GGUF available | ✅ | ✅ (8B and 70B) |
| HuggingFace | ✅ tiiuae/ | ✅ inceptionai/ |
| Cloud inference | Standard providers | Cerebras (2,000 tok/s) |
| Fine-tuning | Standard LoRA/QLoRA | Arabic-specific datasets available |
| Smallest model | H1R 7B (6GB RAM) | 8B Chat (8GB RAM) |
The practical recommendation
Building an Arabic application? Use Jais. No question. The tokenizer efficiency and dialect support make it the only serious choice.
Building a multilingual application with some Arabic? Use Falcon 2 11B. It handles 11 languages reasonably well and Arabic at a basic MSA level.
Need reasoning on budget hardware? Use Falcon H1R 7B regardless of language. Its 88.1% AIME-24 score beats models 7x its size.
Need both Arabic + reasoning? Run both: Jais 2 8B for Arabic tasks, Falcon H1R 7B for reasoning tasks. Both fit on a 16GB machine simultaneously.
FAQ
Is Falcon better than Jais?
For English tasks and reasoning, yes. Falcon H1R-7B has a hybrid SSM+attention architecture that excels at math, logic, and coding. For Arabic tasks, Jais is significantly better — it was built Arabic-first with a custom tokenizer that handles Arabic 2-3x more efficiently than general-purpose models.
Are both free?
Yes. Falcon is Apache 2.0 licensed — fully free for commercial and non-commercial use. Jais has open weights available on HuggingFace. Both can be downloaded and run locally via Ollama at no cost. Falcon H1R-7B runs on as little as 6GB RAM.
Which supports Arabic better?
Jais, by a wide margin. It was specifically trained on 126B Arabic tokens with a custom Arabic-centric tokenizer, and supports Egyptian, Gulf, Levantine, and Moroccan dialects plus Arabic-English code-switching. Falcon handles Modern Standard Arabic at a basic level but lacks dialect support.
Related: What is Falcon? · What is Jais? · How to Run Falcon Locally · How to Run Jais Locally · Yi vs Qwen vs DeepSeek · Sovereign AI Models 2026