Falcon

What is Falcon?

Falcon is the open-source LLM family developed by the Technology Innovation Institute (TII) in Abu Dhabi, first launched in 2023 with Falcon 7B and 40B and now several generations past the Falcon 3 release. Falcon earned its reputation early by briefly topping the Hugging Face Open LLM Leaderboard with Falcon 40B, and TII kept pushing with Falcon 180B and Falcon 2 multimodal in 2024, then Falcon 3 and the state-space Falcon Mamba 7B. The current line-up is anchored by Falcon-H1, a hybrid Transformer plus Mamba architecture shipped in 0.5B, 1.5B, 1.5B-Deep, 3B, 7B, and 34B base and instruct variants, alongside Falcon-E for edge hardware where it runs on CPUs rather than GPUs, Falcon H1R 7B for reasoning, Falcon-H1-Arabic, the very small Falcon-H1-Tiny and Tiny-R checkpoints, and Falcon Perception, a multimodal model that reads and interprets images from natural-language prompts. TII claims Falcon-H1 models outperform models twice their size, with the 0.5B checkpoint landing close to typical 7B models. Falcon models are released under the TII Falcon License, in part based on Apache 2.0 with additional acceptable-use provisions, in practical terms, Falcon is free to download, fine-tune, and use commercially for the vast majority of applications. The widely repeated managed-hosting restriction, which requires a separate TII agreement before offering shared instances as an inference or fine-tuning API, actually lives in the older Falcon 180B license and is not carried into the December 2024 TII Falcon License that covers Falcon 3 and the Falcon-H1 generation. Where Falcon stands out is as a credible non-US, non-Chinese open LLM option: for governments, enterprises, and researchers who want to diversify their LLM supply chain away from Meta, Google, and Alibaba, Falcon offers an independent pretraining lineage with different data sources, languages, and safety tuning. It is particularly strong in Arabic and MENA-region languages where major Western models lag.

Pricing

Weights, Free: All Falcon models can be downloaded from Hugging Face or falconllm.tii.ae under the TII Falcon License. Commercial use is permitted for the vast majority of applications without fees.

Inference providers: Falcon is available on Hugging Face Inference, Amazon Bedrock Marketplace, AWS SageMaker JumpStart, and community-hosted endpoints at standard per-token rates.

Self-hosting: Falcon-E is designed to run on CPUs rather than GPUs, and the small Falcon-H1 checkpoints (0.5B to 3B) plus Falcon-H1-Tiny run comfortably on a laptop. Falcon-H1 7B suits a single consumer GPU; Falcon-H1 34B and Falcon 180B need server-class multi-GPU setups.

Licensing note: The requirement to negotiate a separate TII agreement before offering shared managed instances as an inference or fine-tuning API sits in the Falcon 180B license. The December 2024 TII Falcon License, which covers Falcon 3 and the Falcon-H1 generation, does not include that clause, and Falcon 7B and Falcon 40B remain under plain Apache 2.0.

Key Features

• TII Falcon License, in part based on Apache 2.0, permissive for most commercial use
• Falcon-H1 hybrid Transformer plus Mamba generation: 0.5B, 1.5B, 1.5B-Deep, 3B, 7B, 34B
• Falcon H1R 7B for math, coding, logic, and instruction-following reasoning
• Falcon-E for edge deployment, built to run on CPUs rather than GPUs
• Falcon-H1-Tiny and Tiny-R down to 0.09B for tiny-footprint reasoning
• Falcon Perception, a multimodal model that reads and understands images
• Falcon-H1-Arabic and Falcon Arabic for MENA-region language support
• Earlier generations still published: Falcon 3, Falcon Mamba 7B, Falcon 2, Falcon 180B, Falcon 40B
• Pre-trained on RefinedWeb dataset
• Available on Hugging Face, Amazon Bedrock Marketplace, and AWS SageMaker JumpStart
• Backed by UAE's Technology Innovation Institute

Pros & Cons

FAQ