The Emirates Technology Innovation Institute (TII) has released a new generation of open source small language model Falcon3 series, which includes four models of different sizes and provides two variants: basic version and command version. This series of models performs well on the Hugging Face rankings, outperforming open source models of the same size and even outperforming competitors such as Google, Meta, and Alibaba in multiple benchmark tests. The Falcon3 series is efficient and low-cost, and is particularly suitable for devices and application scenarios with limited computing resources, such as customer service, healthcare, and the Internet of Things. Its training data is large in scale and uses advanced architecture and mechanisms to minimize memory usage and improve inference efficiency. TII also provides the Falcon Playground test environment to facilitate developers and researchers to try it out.
Picture source note: The picture is generated by AI, and the picture authorization service provider Midjourney
The Falcon 3’s performance has topped the Hugging Face rankings, outperforming open source models of the same size, such as Meta’s Llama and Qwen-2.5. In particular, the 7B and 10B versions have demonstrated leading technical advantages in reasoning speed, language understanding, instruction execution, and code and mathematics tasks, and even surpassed competitors such as Google, Meta, and Alibaba in multiple benchmark tests.
Compared with traditional large language models (LLM), SLM models have the advantages of high efficiency and low cost due to their fewer parameters and simpler design, and are especially suitable for applications in customer service, healthcare, Internet of Things and other fields. According to market research firm Values Reports, the SLM market is expected to grow at an average annual rate of 18% over the next five years.
The training data scale of the Falcon3 series reaches 14 trillion tokens, which is more than twice that of its predecessor Falcon2. This series adopts a decoder-only architecture and a grouped query attention mechanism to minimize memory usage while improving inference efficiency. Falcon3 supports four languages, including English, French, Spanish and Portuguese, and is equipped with a 32K context window, which can handle long input text and meet the needs of various industries.
TII said the base model of Falcon3 is suitable for general-purpose tasks, while the command version is optimized for conversational tasks such as customer service and virtual assistants. The launch of this series will further promote the development of edge computing and privacy-sensitive applications, supporting scenarios such as personalized recommendations, data analysis, medical diagnosis, and supply chain optimization.
All Falcon3 models are released under the TII Falcon License 2.0, a permissive license based on Apache 2.0 that supports responsible AI development and deployment. To help developers and researchers get started, TII also launched the Falcon Playground test environment, where users can try out these models before integrating them.
The open source release of the Falcon3 series lowers the threshold for AI technology application, provides developers and researchers with powerful tools, accelerates the application and innovation of AI technology in various fields, and heralds the trend of further popularization and democratization of AI technology.