Falcon LLM

What is Falcon LLM?

Falcon LLM is a family of generative language models from the Technology Innovation Institute (TII) in Abu Dhabi. TII says it has released the models as open source or open access. The lineup spans text, multimodal, reasoning, hybrid-architecture, edge and Arabic models.

Key capabilities of Falcon LLM

• Falcon 3: Multimodal models for text, image, video and audio that can run on light infrastructure.
• Falcon-H1: Hybrid design combining Transformer and Mamba.
• Falcon-E: Edge models aimed at CPU devices.
• Reasoning models: Falcon H1R 7B and tiny reasoning models.
• Arabic models: Falcon-H1-Arabic and Falcon Arabic.
• Large models: Falcon 180B and 40B.

How Falcon LLM works

Developers download the weights, some of which are hosted on Hugging Face, and run them on their own infrastructure. Smaller models such as Falcon 3 and Falcon-E target laptops and CPUs. Each model family has its own license terms.

Who uses Falcon LLM?

Developers and companies building applications, researchers including university researchers through the Falcon LLMs program, and Arabic-speaking users and UAE communities via the Arabic and Emirati models.

Falcon LLM pricing

TII lists no price for Falcon models. Falcon 40B is Apache 2.0, and Falcon 180B uses a royalty-free license based on Apache 2.0, with a separate arrangement required to offer shared hosted inference services. Compute costs are yours.

Falcon LLM alternatives

Preferred Networks is a Japanese AI company that builds its own models and hardware for domestic use. PaddlePaddle is Baidu's deep learning framework rather than a model family. fast.ai provides educational courses and a deep learning library.