Back to All Falcon Models
Falcon 3: Making Advanced AI Accessible and Available to Everyone, Everywhere Experience unmatched performance and scalability on lightweight devices such as laptop and energy constraint infrastructure
Try our Chat
Download Falcon 3 Now
About Falcon 3

Revolutionizing
AI for All

AI continues to redefine industries and transform our interactions with technology, but accessibility remains a critical challenge. Advanced AI models often require robust infrastructure, limiting their reach.

Falcon 3 has been meticulously designed to address this gap.

As an open-source large language model (LLM), Falcon 3 is designed to democratize advanced AI by combining outstanding performance with the ability to run on lightweight devices, including laptops.

Released under TII’s Falcon License 2.0, Falcon 3 is a pioneering step toward making advanced AI tools available to all.

Performance
Benchmark
Efficiency is also a key and Falcon3 is competitive on the generation side. It has a similar behavior to the best-in-class models with an average of 82+ tokens per second for the Falcon3 10B to 244+ tokens per second for the Falcon3 1B on an H100. This performance is recovered with no optimization on a TGI API. Falcon 3 is a performance leader in the realm of small LLMs, surpassing global benchmarks and rivaling models from leading institutions. It ranks first on Hugging Face leaderboards, outperforming other open-source models like Meta’s Llama variants and setting new thresholds for excellence in its size category. Falcon 3 Base model surpasses the performance threshold set by Qwen models, demonstrating robust foundational capabilities. Meanwhile, the Instruct/Chat model rank first globally, showcasing remarkable fine-tunability and leading performance in conversational and task-specific benchmarks. Falcon 3’s innovative design, from its scalable model sizes to its resource-efficient deployment, ensures it is well-positioned to serve the needs of diverse users while setting new benchmarks in the AI landscape. The improvements aren’t just theoretical, Falcon 3 delivers real-world impact across industries, empowering users to achieve more with less infrastructure.
Our Ambitions for Falcon 3
Democratized AI Access Falcon 3 by TII offers models that are small, efficient, and capable of running on lightweight infrastructures. It ensures high performance without requiring extensive computational resources.
High Accessibility & Performance Designed for developers, researchers, and businesses, Falcon 3 empowers users to leverage cutting-edge AI tools while maintaining ease of use and accessibility.
Improved Efficiency & Fine-Tuning Falcon 3 builds on the success of Falcon 2, delivering enhanced reasoning, fine-tuning capabilities, and improved efficiency across a wide range of use cases.
Commitment to Innovation Reinforcing Technology Innovation Institutes’s (TII) mission, Falcon 3 fosters inclusive, open-source innovation, providing the global community with state-of-the-art AI models.
Model Architecture
Optimized Decoder-Only Design

Falcon 3’s architecture is based on a decoder-only design using flash attention 2 to grouped query attention. It integrates Grouped Query Attention (GQA) to share parameters, minimizing memory for Key-Value (KV) cache during inference, ensuring faster and more efficient operations.

Advanced Tokenization

With a tokenizer supporting a high vocabulary of 131K tokens—double that of Falcon 2—Falcon 3 offers superior compression and improved downstream performance, enhancing its ability to handle diverse tasks.

Enhanced Long-Context Training

Trained natively with a 32K context size, Falcon 3 demonstrates exceptional long-context capabilities, delivering enhanced performance for extended input data compared to its predecessors.

The Falcon 3 series represents a huge leap forward in AI technology. Trained on an impressive 14 Trillions tokens, Falcon 3 more than doubles the capacity of its predecessor, Falcon 180B, ensuring a significant boost in performance and capability. The initial training was followed by multiple stages to improve reasoning and math performance with high-quality data and context extension with natively long context data. Falcon 3 was trained on 4 main languages (English, Spanish, Portuguese and French) to ensure a much higher, earning capability and quality for those languages.
Our Approach to Responsible AI
Falcon 3 is released under the TII Falcon License. This framework promotes the responsible development and deployment of AI while empowering the global community to innovate freely. By emphasizing ethical AI practices, Falcon 3 balances openness with accountability, ensuring technology is used for the benefit of society.
Building with Falcon 3

Advanced AI for Everyone, Everywhere

Falcon’s quantized versions, such as GGUF, AWQ, and GPTQ (in int4, int8, and 1.58 Bitnet), make it highly efficient, even for resource-constrained environments. Optimized for lightweight systems, Falcon 3 is a game-changer. The model family comprises four text-based models—1B, 3B, 7B, and 10B—with both Base and Instruct versions tailored for different needs. Falcon 3 models can be further customized through tools like vLLM, Llama.cpp, and MLX, ensuring seamless adoption for developers.

These innovations reflect our commitment to ensuring AI is accessible and efficient for a wide range of users.

Falcon 3 is versatile, designed for both general-purpose and specialized tasks, providing immense flexibility to users. 

Its Base model is perfect for generative applications, while the Instruct model excels in conversational tasks like customer service or virtual assistants.

Falcon 3 is straight forward to implement, whether you’re a startup seeking to enhance user experience or a researcher exploring innovative AI applications. For organizations and individuals with limited computational resources, Falcon 3’s quantized versions offer rapid deployment and optimized efficiency without compromising performance.

What’s Next?

TII has ambitious plans for Falcon 3 with new capabilities currently in the work, further broadening its applicability.

This phased approach ensures users can fully explore and adapt the model’s capabilities, paving the way for greater adoption and impact.

Try our Chat
Download Falcon 3 Now