Falcon Models

A 7-billion-parameter language model purpose-built for the vocabulary, expressions, idioms and cultural references that shape how Emirati communities communicate.

Falcon-ASR is a compact 1.6-billion-parameter Automatic Speech Recognition model designed to convert spoken language into written text.

Falcon-OCR-Arabic is a 270-million-parameter model built for Arabic optical character recognition and document understanding, extracting text and structured information from images and documents.

Falcon Perception is a next-generation multimodal AI model that enables systems to see, read, and understand images using natural language prompts. By combining vision and language capabilities in a single architecture, Falcon Perception simplifies how AI interprets visual information while remaining efficient.

The Falcon‑H1‑Tiny series demonstrates that intelligence can indeed fit into small models, with specialized variants delivering multilingual understanding, strong instruction following, robust tool‑use and coding capabilities. Falcon‑H1‑Tiny‑R series redefines reasoning at scale by matching the performance of significantly larger reasoning models. Falcon‑H1‑Tiny‑R was trained exclusively on reasoning data and applied to compact architectures, unlocking powerful reasoning performance in the 0.6B and 0.09B configurations.

Falcon-H1-Arabic is a complete departure from traditional large language models.

Built on a new hybrid Mamba-Transformer architecture, Falcon-H1-Arabic represents a significant evolution of the Falcon H1 family, combining higher accuracy with greater efficiency.

Advanced reasoning capabilities in mathematics, coding, logic and instruction-following in a highly efficient 7B model.

Built for the Modern Standard Arabic and dialects, Falcon Arabic brings advanced language AI to the Arabic-speaking world.

AI innovation – and its deployment in the real world – is speeding up. Until now, users had to make a trade-off between performance and efficiency. The unique hybrid architecture of Falcon-H1 (based on Transformer and Mamba architecture) gives you a family of AI models that delivers the best of both worlds – combining strong general-purpose comprehension with highly efficient processing. Thanks to an updated version of Maximal Update Parametrization it is possible to train the models reliably at very large scales; in fact, it makes the process of increasing model size much safer and more efficient.

Falcon Edge redefines what’s possible with small-scale language models. As AI moves beyond the cloud and into real-world devices, Falcon-E is built to work where infrastructure is limited, running fast and efficiently on CPUs, not GPUs.

The latest updates to the Falcon 3 series introduce exceptional multimodal functionalities, enabling the models to process not only text but also images, and – for the first time in the Falcon series – video and audio data. Falcon 3 can run on light infrastructures, even laptops, without sacrificing performance. The Falcon 3 ecosystem contains four scalable models tailored for diverse applications. Falcon 3 supports several languages and is optimized for resource efficiency. This latest iteration of TII’s open-source large language model series aims to democratize access to AI.

Falcon Mamba 7B is the first open source released State Space Language Model (SSLM), a new revolutionary architecture for Falcon models. Falcon Mamba 7B is the no. 1 globally performing open source SSLM in the world, as independently verified by Hugging Face. SSLMs have a low memory cost and don’t require additional memory to generate arbitrary long blocks of text. Falcon Mamba 7B also outperforms traditional transformer architecture models such as Meta’s Llama 3.1 8B and Mistral’s 7B. New model reflects the innovation and pioneering approach of Abu Dhabi in AI research and development.

Open-Source, Multilingual, and Multimodal – and is only AI Model with Vision-to-Language Capabilities. New Falcon 2 11B Outperforms Meta’s Llama 3 8B, and Performs on par with leading Google Gemma 7B Model, as Independently Verified by Hugging Face Leaderboard. Next up, we’re looking to add 'Mixture of Experts' to enhance Falcon 2’s capabilities even further.

Falcon-7B is a 7B parameters causal decoder-only model built by TII and trained on 1,500B tokens of RefinedWeb enhanced with curated corpora. It is made available under the Apache 2.0 license.

Falcon 40B was the world’s top-ranked open-source AI model when launched. Falcon has 40 billion parameters and was trained on one trillion tokens. For two months following its launch, Falcon 40B ranked #1 on Hugging Face’s leaderboard for open source large language models (LLMs). Offered completely royalty-free with weights, Falcon 40B is revolutionary and helps democratize AI and make it a more inclusive technology.

The multilingual Falcon 40B LLM works well with English, German, Spanish, French, Italian, Portuguese, Polish, Dutch, Romanian, Czech, and Swedish languages. The foundational LLM serves as a versatile base model that can be fine-tuned for specific requirements or objectives.

Falcon 40B launched a Call for Proposals from scientists, researchers, and innovators for inspiring use cases and applications with the most exceptional use cases to receive an investment of training computing power to work on the powerful model to shape transformative solutions. The model uses only 75 percent of GPT-3’s training compute, 40 percent of Chinchilla AI’s, and 80 percent of PaLM-62B’s.

One of the core differences in the development of Falcon was the quality of the training data. The size of the pre-training data collected for Falcon 40B was nearly five trillion tokens gathered from public web crawls (~80%), research papers, legal text, news, literature, and social media conversations.

Since LLMs are particularly sensitive to the data they are trained on, our team built a custom data pipeline to extract high-quality pre-training data using extensive filtering and deduplication, implemented both at the sample level and at the string level.

Falcon 180B is a super-powerful language model with 180 billion parameters, trained on 3.5 trillion tokens. It's currently at the top of the Hugging Face Leaderboard for pre-trained Open Large Language Models and is available for both research and commercial use..

This model performs exceptionally well in various tasks like reasoning, coding, proficiency, and knowledge tests, even beating competitors like Meta's LLaMA 2.

Among closed source models, it ranks just behind OpenAI's GPT 4, and performs on par with Google's PaLM 2 Large, which powers Bard, despite being half the size of the model.