"Technological innovation" “Falcon Inference” is revealed as the world’s best 7B artificial intelligence model

ABU DHABI, 5 JANUARY / WAM / The Technology Innovation Institute, the applied research arm of the Advanced Technology Research Council in Abu Dhabi, announced the launch of the Falcon H1R 7B, a new generation artificial intelligence model, which constitutes a qualitative step towards making advanced artificial intelligence available on a broader scale, through world-class reasoning capabilities within a compact, highly efficient model that is openly available to everyone.

The model, which includes only 7 billion parameters, is distinguished by its ability to compete with and exceed the performance of larger open source artificial intelligence models from around the world, including models developed by Microsoft (Phi 4 Reasoning Plus 14B), Alibaba (Qwen3 32B), and Nvidia (Nemotron H 47B).

This launch embodies the Institute’s commitment to pushing the boundaries of innovation in highly efficient artificial intelligence, and reflects the country’s growing presence in leading technological development at the global level.

His Excellency Faisal Abdulaziz Al Bannai, Advisor to the President of the UAE for Strategic and Advanced Technology Research, Secretary-General of the Advanced Technology Research Council, said that the Falcon H1R embodies the UAE’s vision of developing open and responsible artificial intelligence that achieves a tangible impact locally and globally.

He added that providing advanced reasoning capabilities within an effective and small-scale model contributes to expanding the benefit of artificial intelligence, supporting economic growth, enhancing the status of scientific research, and supporting sustainability and technological flexibility in the long term.

The Falcon H1R 7B is based on the Falcon H1-7B model, introducing an advanced training method and a Transformer-Mamba hybrid architecture, which enhances accuracy levels and increases response speed.

For her part, Dr. Najwa Al-Araj, CEO of the Technology Innovation Institute, said that the Falcon H1R 7B embodies a remarkable development in the inference capabilities of small-sized artificial intelligence models, as it achieves near-perfect results in high-level standard tests, with exceptional efficiency in memory and energy consumption, which supports its readiness for real-world applications and enhances sustainability.”

This trend opens the horizons of what is known as “latent intelligence,” allowing the model to reason more effectively and efficiently.

The Falcon H1R 7B also sets new limits on what is known as the Pareto limit, where the optimal balance between performance and speed is achieved without compromising on quality.

The Falcon H1R 7B performed outstandingly across a range of competitive benchmark tests, including the following: Math: It scored 88.1% on the AIME-24 test, besting the April 1.5 (15B) ServiceNow AI model which achieved 86.2%, demonstrating the ability of a compact 7B model to compete with or even surpass much larger systems.

Programming and proxy tasks: It achieved an accuracy of 68.6%, the best performance in its class among models under 8B, and it also scored higher results in the LCB v6, SciCode Sub, and TB Hard tests, where the Falcon H1R’s performance was 34% compared to 26.9% for the Chinese model DeepSeek R1-0528 Qwen 3 8B, also exceeding larger models such as the Qwen3-32B, which scored 33.4%.

General Reasoning: Demonstrated strong reasoning and instruction-following abilities, achieving similar or close performance to larger models such as Microsoft’s Phi 4 Reasoning Plus (14B), while using only half the number of operands.

Efficiency: Reach speeds of up to 1,500 codes/s/GPU at Batch 64, nearly twice as fast as the Qwen3-8B, thanks to its hybrid Transformer–Mamba architecture, delivering faster, scalable performance without compromising on accuracy.

For his part, Dr. Hakim Hasid, chief researcher at the Artificial Intelligence and Digital Sciences Research Center of the Institute of Technology Innovation, said that this model represents the result of world-class research and engineering, and embodies how scientific rigor and scalable design can go hand in hand, expressing pride in presenting a model that enables the community of researchers and developers to build smarter, faster, and more accessible artificial intelligence systems.

In keeping with ITI’s commitment to transparency and collaboration in the field of artificial intelligence, the Falcon H1R 7B has been released as an open source model under ITI’s Falcon License.

Developers, researchers, and organizations around the world can access the model via the Hugging Face platform, along with a comprehensive technical report reviewing training strategies and the model’s performance on the most prominent benchmark inference tests.

This new release builds on the continued global successes of the Institute of Technology Innovation’s Falcon program.

Since its launch, Falcon models have continued to be among the best-performing artificial intelligence systems in the world, with the first four generations achieving first place globally in their categories.

Throughout successive development phases, Falcon has set new standards in performance, efficiency and practicality, ensuring that compact Sovereign models can outperform much larger systems.

Together, these stations confirm the growing leadership of Abu Dhabi and the UAE in advanced artificial intelligence technologies, and the ability of the Institute of Technology Innovation to provide globally competitive research.

• For more: Follow Khaleejion 24 Arabic, Khaleejion 24 English, Khaleejion 24 Live, and for social media follow us on Facebook and Twitter