Arabic Excellence

Why Native Matters

Most "Arabic-capable" LLMs are English-first models that treat Arabic as a second-class citizen. They rely on translation layers or inefficient tokenization that fragments Arabic words, leading to:

• Higher Latency: More tokens per word means slower generation.
• Increased Cost: You pay per token; inefficient tokenization doubles your bill.
• Cultural Hallucinations: Direct translation misses idiom, context, and nuance.

The Tokenization Advantage

We rebuilt the tokenizer vocabulary, adding 20,000+ native Arabic tokens. This allows our models to process whole words and phrases rather than fragmented characters.

Dialect Support

Modern Standard Arabic (MSA) is the language of business, but your users speak dialects. MX4 Atlas is fine-tuned on a massive corpus of dialectal data.

Performance by Dialect

Accuracy and latency metrics across all supported Arabic dialects:

Benchmarks

We evaluate our models on standard Arabic benchmarks (OpenArabic, ALDi) and our own proprietary sovereign evaluation set.

Fine-Tuning by Dialect

Recommendations for fine-tuning on specific Arabic dialects:

Cost Efficiency

Native Arabic tokenization dramatically reduces costs compared to translation-based approaches:

1# Cost comparison: 1000 character Arabic text
2text = "المملكة العربية السعودية هي أكبر منتج نفط في العالم"
3
4# Standard models (4.2 tokens per word)
5standard_tokens = len(text.split()) * 4.2  # ≈ 63 tokens
6
7# MX4 Atlas (1.6 tokens per word) 
8atlas_tokens = len(text.split()) * 1.6  # ≈ 24 tokens
9
10# Cost saving with Atlas
11standard_cost = (standard_tokens / 1000000) * 0.15  # $0.15/1M tokens
12atlas_cost = (atlas_tokens / 1000000) * 0.15
13
14savings = ((standard_cost - atlas_cost) / standard_cost) * 100
15print(f"Cost reduction: {savings:.1f}% with Atlas")  # ~62% savings