FalconUnclaimed

Falcon LLM is a family of generative large language models built by the Technology Innovation Institute (TII), a research centre that is the applied research pillar of Abu Dhabi's Advanced Technology Research Council. The page presents the models as a suite covering several uses: multimodal image understanding, compact reasoning, Arabic language work, hybrid architectures for efficient processing, and small models that run on CPUs.

The lineup on the site includes Falcon Perception, a multimodal model that reads and interprets images from natural language prompts; Falcon H1R 7B, a 7-billion-parameter model for reasoning in mathematics, coding, logic and instruction-following; Falcon-H1, a hybrid model based on Transformer and Mamba architectures; Falcon-H1-Arabic and Falcon Arabic, built for Modern Standard Arabic and dialects; Falcon-E, designed to run on CPUs where infrastructure is limited; and Falcon 3, which adds image, video and audio processing. The site also lists Falcon-Emirati, Falcon-ASR and Falcon-OCR-Arabic as new models for Emirati vocabulary, speech transcription and text extraction from images.

The page says the team has open sourced or made open access all its Falcon models, and that Falcon 180B was released for research and commercial use. Falcon Mamba 7B is described as an open source State Space Language Model. The site also mentions a Falcon Foundation, a non-profit for open-source generative AI, and calls for proposals in which developers and university researchers can apply for compute to build ideas with the models.

The page describes a dataset called REFINEDWEB and a licensing section that sets out the terms for downloading, using and deploying each model. It does not state a price for any model, so this listing gives no cost figures.

The models are aimed at developers, researchers and organisations that want to build AI applications, with particular attention to Arabic-speaking users and to settings with limited infrastructure.