About
Falcon-H1-Arabic
Falcon-H1-Arabic is a complete departure from traditional large language models.
Built on a new hybrid Mamba-Transformer architecture, Falcon-H1-Arabic represents a significant evolution of the Falcon H1 family, combining higher accuracy with greater efficiency.
It was designed specifically to deliver state-of-the-art Arabic language understanding, reasoning, and long-context performance.
Falcon-H1- Arabic sets a new global benchmark for Arabic AI, achieving leading performance on the Open Arabic LLM Leaderboard (OALL).
Proof Points:
Purpose-built Arabic AI designed specifically for high-quality Arabic language understanding
The first Arabic LLM with a Mamba-Transformer hybrid architecture for higher accuracy and efficiency
Trained on native-Arabic datasets without reliance on machine translation
Supports Modern Standard Arabic and dialects, including Gulf, Levantine, North African and Egyptian
Leading global open-source AI model
Multingual model supporting 18 languages
Available in flexible model sizes for varied infrastructure and compute environments: 3B, 7B and 34B.
Supports context windows of up to 256k tokens equivalent to ~ 100k words
Performance Benchmark
We purpose-built Falcon-H1-Arabic to support the linguistic and cultural complexity of Arabic across real-world applications.
Major Improvements In:
The Falcon-H1-Arabic Family is available in 3B, 7B, and 34B parameter variants for flexibility across a wide range of infrastructure and deployment needs.
They may be small, but they outperform the mighty:
3B model / 01
Outperforms the best state-of-the-art models with more than 10 points on OALL benchmark. The model is suitable for light-weight applications and agentic AI.
7B model / 02
Surpasses all <10B models on OALL benchmarks. Provides a good tradeoff between efficiency and accuracy
34B model / 03
Outperforms even 70B+ parameter systems on Arabic language benchmarks
High-quality Arabic AI doesn’t need excessive model scale or infrastructure overhead
For
Long
Context
Falcon-H1-Arabic supports context windows of up to 256K tokens, processing and reasoning across large volumes of Arabic text in a single interaction. Users can analyze lengthy legal documents, medical records, academic papers, or enterprise knowledge bases without losing coherence or continuity — a capability previously unavailable at this level for Arabic language models.