The Technology Innovation Institute (TII) in Abu Dhabi has released the next-generation open-source AI model, Falcon3. With a training volume of 140 trillion tokens and an optimized architectural design, it sets a new record for performance on consumer-grade hardware. This training scale is more than double that of its predecessor, Falcon2.
The Falcon3 series offers four specifications: 1B, 3B, 7B, and 10B, each available in a base version and an Instruct version optimized for dialogue. Although specific versions are provided for English, French, Spanish, and Portuguese, all models can handle most commonly used languages.
In third-party language model evaluations on Hugging Face, Falcon3 outperformed mainstream open-source models, including Meta's Llama-3.1-8B, Qwen2.5-7B, Mistral's NeMo-12B, and Google's Gemma2-9B, demonstrating strong competitiveness.
Falcon3 performs better in relevant benchmark tests compared to similarly scaled competitors like Mistral, Alibaba, Meta, and Google. | Image: Technology Innovation Institute
TII emphasizes the model's ease of use, ensuring compatibility with standard APIs and libraries, and providing resource-optimized quantized versions for specific hardware configurations. Additionally, the institute has launched a free chatbot for users to test and provide feedback. The product's interface design draws on the successful experience of ChatGPT, incorporating similar features like project folders.
Looking ahead, TII plans to expand the capabilities of the Falcon3 series in early 2025, introducing multimodal models that support image, video, and voice processing. Currently, all models are available for free download on the Hugging Face platform under the TII Falcon license based on Apache 2.0, which includes guidelines for promoting responsible AI use.
Falcon Chat has a very similar interface to ChatGPT and includes comparable features. | Image: Screenshot from THE DECODER
This release marks another significant advancement in the open-source AI field, particularly achieving breakthrough accomplishments in enhancing AI performance on consumer-grade hardware. With the addition of multimodal capabilities, Falcon3 is expected to bring more innovative application possibilities to the AI open-source community in 2025.