TII Next-Generation AI Model Outperforms Larger Rivals
KUALA LUMPUR, Jan 6 (Bernama) -- The Technology Innovation Institute (TII) has released the Falcon H1R 7B, a next-generation artificial intelligence (AI) model designed to make advanced reasoning AI more accessible through a compact, efficient, and openly available format.
With just seven billion parameters, Falcon H1R 7B outperforms several larger open-source AI models from global technology players, including Microsoft, Alibaba, and NVIDIA, reinforcing TII’s position at the forefront of efficient AI innovation and underscoring the United Arab Emirates’ growing influence in global technology leadership.
TII Chief Executive Officer, Dr Najwa Aaraj said Falcon H1R 7B represents a major advance in compact AI reasoning, achieving near-perfect scores on elite benchmarks while maintaining exceptionally low memory and energy requirements, making it suitable for real-world deployment.
According to TII in a statement, the model builds on the Falcon H1-7B foundation through a specialised training approach and a hybrid Transformer–Mamba architecture that improves accuracy and speed.
In competitive benchmarks, Falcon H1R 7B delivered 68.6 per cent accuracy, the best performance among models under eight billion (8B) parameters, while demonstrating strong logic and instruction-following capabilities comparable to significantly larger systems.
The model also achieved processing speeds of up to 1,500 tokens per second per GPU at batch 64, nearly doubling the performance of China’s Qwen3-8B, enabled by its hybrid architecture that delivers scalable speed without sacrificing accuracy.
Falcon H1R 7B is released as an open-source model under the Falcon TII License, with global access available via Hugging Face alongside a detailed technical report on training methods and benchmark performance.
The release builds on the global success of TII’s Falcon programme, whose models have consistently ranked among the world’s top-performing AI systems, demonstrating that compact, sovereign models can rival and exceed much larger alternatives.
-- BERNAMA