Falcon 3 is designed to democratize access to high-performance AI, offering models that are both powerful and efficient. Trained on 14 trillion tokens—more than double its predecessor's 5.5 trillion—Falcon 3 demonstrates superior performance across various benchmarks. Notably, it ranks among the top models globally that can operate on a single GPU. Upon its release, Falcon 3 achieved the number one position on Hugging Face's global third-party LLM leaderboard, surpassing other open-source models of similar size, including Meta's Llama variants. In particular, the Falcon 3-10B model leads its category, outperforming all models under 13 billion parameters.

His Excellency Faisal Al Bannai, Secretary General of ATRC and Adviser to the UAE President for Strategic Research and Advanced Technology Affairs, said: “The transformative power of AI is undeniable. Today, we advance our contributions to the AI community, particularly the open-source sector, with the release of the Falcon 3 family of text models. This launch builds upon the foundation we established with Falcon 2, marking a significant step toward a new generation of AI models. Our ongoing commitment to ensuring these powerful tools remain accessible to everyone, everywhere, reflects our dedication to global equity and inclusive innovation.”

Falcon 3 Family

The Falcon 3 series include four model sizes, Falcon3-1B, -3B, -7B and -10B. Compared to its predecessors, significant attention has been placed on seamless integration. The models are fully compatible with widely used APIs (Application Programming Interfaces) and libraries, considerably reducing integration efforts and ensuring ease of use. This ultimately enables users to select a solution that best fits their needs. With Falcon 3 delivering exceptional performance across reasoning, language understanding, instruction following, code generation, and mathematical tasks, it is poised to set new standards in AI capabilities.

The Falcon 3 small models each have a Base and an Instruct variation that each rank among the most powerful in the world for its size. The Base model allows general-purpose generative tasks, whereas the Instruct is a fine-tuned variant for conversational applications. Falcon 3 is available in English, French, Spanish and Portuguese. Falcon 3 models also have quantized versions that enable optimized integration into specialized architectures, being resource efficient and lightweight for rapid deployment and inference.

Dr. Najwa Aaraj, TII’s CEO, said: "Our dedication to pioneering research and attracting top-tier talent has culminated in the development of Falcon 3. The result is a model that exemplifies our pursuit of scientific excellence, offering enhanced efficiency and setting new benchmarks in AI technology."

Dr. Hakim Hacid, Chief Researcher of the TII’s AI and Digital Science Research Center (AIDRC), said: “AI is fast evolving, and we are glad to be an active part of this journey. Falcon 3 pushes the boundaries of small LLMs further, contributing to the open-source community by providing access to a better-performing AI. We are confident that this latest release will open an unlimited range of opportunities and will have tremendous benefits, empowering businesses and individuals to use AI in ways that were previously out of reach.”

Falcon 3 is available immediately for download on HuggingFace and at FalconLLM.TII.ae, along with details on the benchmarks.

TII is also introducing Falcon Playground, a testing environment for end-users, programmers, coders, and researchers to explore Falcon 3 before its official release, providing an opportunity to experiment and offer feedback.

Falcon 3 is licensed under TII Falcon License, the permissive Apache 2.0-based software license that includes an acceptable use policy which promotes the responsible use of AI.

Early January 2025, the Falcon 3 Family of models will introduce new members with an emphasis on multimodal functionalities—including text, images, video, and voice mode.