Some links in this article are affiliate links. We earn a commission at no extra cost to you when you purchase through them. Full disclosure.

Jais 2 is the world’s best Arabic LLM. The 8B version runs on consumer hardware, giving you native Arabic AI for free. Here’s the setup.

Available models

Model	Size	RAM needed	Best for
Jais 2 8B Chat	~5 GB	8 GB	Arabic chat, trained on 126B Arabic tokens
Jais 2 70B Chat	~45 GB	48 GB+	Best Arabic quality, GGUF available
Jais 13B (v1)	~8 GB	12 GB	Legacy, still solid

Setup with Ollama

# Install Ollama
brew install ollama

# Pull Jais (check Ollama library for latest)
ollama pull jais

# Or use GGUF from HuggingFace
# Download from huggingface.co/inceptionai/Jais-2-8B-Chat-GGUF

Setup with HuggingFace

For the full 70B model or more control:

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "inceptionai/Jais-2-8B-Chat"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name, device_map="auto")

# Chat in Arabic
prompt = "اشرح لي كيف يعمل الذكاء الاصطناعي بطريقة بسيطة"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=500)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Setup with llama.cpp (GGUF)

Official GGUF quantizations are available from inceptionai/Jais-2-8B-Chat-GGUF on HuggingFace:

# Download GGUF directly
huggingface-cli download inceptionai/Jais-2-8B-Chat-GGUF --local-dir ./jais-gguf

# Run with llama.cpp
./llama-cli -m ./jais-gguf/jais-2-8b-chat-q5_k_m.gguf -p "اشرح الذكاء الاصطناعي" -n 500

Available quantizations:

Quantization	Size	RAM needed	Arabic quality
Q8_0	~8.5 GB	10 GB	Best
Q5_K_M	~5.5 GB	8 GB	Very good (recommended)
Q4_K_M	~4.5 GB	6 GB	Good
Q3_K_M	~3.5 GB	5 GB	Noticeable loss on dialects

Recommendation: Q5_K_M is the sweet spot. Arabic text is more sensitive to quantization than English because of the complex morphology — going below Q4 causes noticeable quality drops on dialectal Arabic.

Hardware requirements

Hardware	Jais 2 8B	Jais 2 70B
MacBook Air M2 8GB	✅ ~15 tok/s	❌
MacBook Pro M3 16GB	✅ ~25 tok/s	❌
Mac Mini M4 Pro 48GB	✅ ~30 tok/s	✅ ~8 tok/s
RTX 4090 24GB	✅ ~35 tok/s	❌ (VRAM)
A100 80GB	✅ ~50 tok/s	✅ ~20 tok/s

The 8B model is the practical choice for most developers. The 70B model needs serious hardware or a cloud GPU (RunPod, Vultr).

Use cases for developers

Arabic customer support bot

system_prompt = """أنت مساعد خدمة عملاء ودود. أجب على أسئلة العملاء باللغة العربية بطريقة مهنية ومفيدة."""

response = chat(system_prompt, user_message)

Arabic code documentation

prompt = "اكتب توثيقاً بالعربية لهذه الدالة:\n\ndef calculate_tax(income, rate):\n    return income * rate"

Arabic-English translation with context

Jais understands code-switching (mixing Arabic and English), which is common in tech contexts across the Middle East.

Jais vs general-purpose models for Arabic

Model	Arabic quality	Arabic dialects	English	Run locally
Jais 2 8B	✅ Best	✅ MSA + dialects	Good	✅ 8GB
Jais 2 70B	✅ Best	✅ MSA + dialects	Good	48GB+
GPT-5	Good	Basic MSA only	✅ Best	❌ API only
Qwen 3.5	Decent	Basic	✅ Excellent	✅ 20GB
Llama 4	Decent	Basic	✅ Excellent	✅ varies

General-purpose models handle Modern Standard Arabic (MSA) reasonably well, but they struggle with dialects (Egyptian, Gulf, Levantine, Moroccan). Jais was specifically trained on dialectal Arabic data.

Jais + Falcon: the UAE AI stack

For developers building Arabic applications, the optimal stack is:

• Jais 2 for Arabic language tasks (chat, content, translation)
• Falcon H1R for reasoning and coding tasks
• Falcon 2 for multilingual tasks

All open source, all from the UAE ecosystem. See our Falcon vs Jais comparison for detailed analysis.

Related: What is Jais? · Falcon vs Jais · What is Falcon? · Ollama Complete Guide · Best Open Source Coding Models · Self-Hosted AI for Enterprise