--- title: README emoji: πŸš€ colorFrom: red colorTo: indigo sdk: static pinned: false --- **Do you believe in a better tomorrow? We do. Our team of expert researchers live the dream and work to build it every day.** * πŸ¦…πŸ¦… **The second generation of Falcon models has been released open-access, featuring [Falcon2-11B](https://huggingface.co/tiiuae/falcon-11B) and [Falcon2-11B-vlm](https://huggingface.co/tiiuae/falcon-11B-vlm).** * πŸ”₯ **[Falcon-180B](https://huggingface.co/tiiuae/falcon-180b) is now available in open-access! [Try it now in our chat demo!](https://huggingface.co/spaces/tiiuae/falcon-180b-demo)** # News * 🐍 **[FalconMamba-7B](https://huggingface.co/tiiuae/falcon-mamba-7b) is now available.** The first pure SSM model of the Falcon series released under the same permissive license. You can interact with it [here](https://huggingface.co/spaces/tiiuae/falcon-mamba-playground), and find the FalconMamba blogpost **[here](https://huggingface.co/blog/falconmamba)**. * πŸ“Έ **[Falcon2-11B-vlm](https://huggingface.co/tiiuae/falcon-11B-vlm) is now available.** Built on top of the Falcon2-11B model, and released under the same permissive license, this open source model allows users to interact with image content via text. * πŸŽ‰ **TII has just released a new generation of models, starting with [Falcon2-11B](https://huggingface.co/tiiuae/falcon-11B)**, a 11B parameters causal decoder-only model and trained over 5,000B tokens of [RefinedWeb](https://huggingface.co/datasets/tiiuae/falcon-refinedweb) enhanced with curated corpora. The model is made available under the [TII Falcon License 2.0](https://falconllm-staging.tii.ae/falcon-2-terms-and-conditions.html), the permissive Apache 2.0-based software license which includes an [acceptable use policy](https://falconllm-staging.tii.ae/falcon-2-acceptable-use-policy.html) that promotes the responsible use of AI. * πŸ’₯ **TII has open-sourced Falcon-180B for research and commercial utilization!** Access the [180B](https://huggingface.co/tiiuae/falcon-180b), as well as [7B](https://huggingface.co/tiiuae/falcon-7b)/[40B](https://huggingface.co/tiiuae/falcon-40b) models, and explore our high-quality web dataset, [RefinedWeb](https://huggingface.co/datasets/tiiuae/falcon-refinedweb). * ✨ **Falcon-[40B](https://huggingface.co/tiiuae/falcon-40b)/[7B](https://huggingface.co/tiiuae/falcon-7b) are now available under the Apache 2.0 license**, TII has [waived all royalties and commercial usage restrictions](https://www.tii.ae/news/uaes-falcon-40b-worlds-top-ranked-ai-model-technology-innovation-institute-now-royalty-free). # FalconMamba LLM We are excited to announce the release of our groundbreaking LLM model with a pure SSM architecture, setting a new benchmark by outperforming all previous SSM models and achieving performance on par with leading transformer-based models. More details on the new models and their performance can also be found in our [FalconMamba blogpost](https://huggingface.co/blog/falconmamba). | **Artefact** | **Link** | **Type** | **Details** | |---------------------|------------------------------------------------------------------|-------------------------|-------------------------------------------------------------------| | 🐍 **FalconMamba-7B** | [Here](https://huggingface.co/tiiuae/falcon-mamba-7b) | *pretrained model* | 7B parameters pure SSM trained on ~6,000 billion tokens. | | FalconMamba-7B-Instruct | [Here](https://huggingface.co/tiiuae/falcon-mamba-7b-instruct) | *instruction/chat model* | Falcon-Mamba-7B finetuned using only SFT.| | FalconMamba-7B-4bit | [Here](https://huggingface.co/tiiuae/falcon-mamba-7b-4bit) | *pretrained model* | 4bit quantized version using GGUF.| | FalconMamba-7B-Instruct-4bit | [Here](https://huggingface.co/tiiuae/falcon-mamba-7b-instruct-4bit) | *instruction/chat model* | 4bit quantized version using GGUF.| # Falcon2 LLM Falcon2 LLM is TII's new flagship series of large language models, where we focused on building smaller models with enhanced performance to enable cheaper inference that can encourage the development of more downstream applications and improve the general usability of our models. Papers: - [Falcon2-11B Technical Report, Malartic et al. 2024](https://www.arxiv.org/abs/2407.14885) More details on the new models and their performance can also be found in our [Falcon2 blogpost](https://huggingface.co/blog/falcon2-11b). See below for a detailed list of artefacts in the Falcon2 LLM family: | **Artefact** | **Link** | **Type** | **Details** | |---------------------|------------------------------------------------------------------|-------------------------|-------------------------------------------------------------------| | πŸ¦…πŸ¦… **Falcon-11B** | [Here](https://huggingface.co/tiiuae/falcon-11B) | *pretrained model* | 11B parameters trained on over 5000 billion tokens. | | πŸ¦…πŸ“Έ **Falcon-11B-vlm** | [Here](https://huggingface.co/tiiuae/falcon-11B-vlm) | *vision adapted model* | Integrating the pretrained CLIP ViT-L/14 vision encoder with our Falcon2-11B chat-finetuned model, and trained with image-text data. | # Falcon LLM Falcon LLM is TII's flagship series of large language models, built from scratch using a custom data pipeline and distributed training library [Almazrouei et al. 2023](https://arxiv.org/abs/2311.16867). Papers: - [RefinedWeb, Penedo et al. 2023](https://proceedings.neurips.cc/paper_files/paper/2023/hash/fa3ed726cc5073b9c31e3e49a807789c-Abstract-Datasets_and_Benchmarks.html) - [The Falcon Series of Open Language Models, Almazrouei et al. 2023](https://arxiv.org/abs/2311.16867) To promote collaborations and drive innovation, we have open-sourced a number of artefacts: * The **Falcon-180B** pretrained and chat models, under the [Falcon-180B TII license](https://huggingface.co/spaces/tiiuae/falcon-180b-license/blob/main/LICENSE.txt). Falcon-180B is the largest and most powerful open-access model available. * The **Falcon-7/40B** pretrained and instruct models, under the Apache 2.0 software license . Falcon-7B/40B models are state-of-the-art for their size, outperforming other open-source models on NLP benchmarks. * The **RefinedWeb** dataset, a massive web dataset with stringent filtering and large-scale deduplication, enabling models trained on web data alone to match or outperform models trained on curated corpora. See πŸ““ [the paper](https://arxiv.org/abs/2306.01116) for more information. RefinedWeb is licensed under ODC-By 1.0. See below for a detailed list of artefacts in the Falcon LLM family: | **Artefact** | **Link** | **Type** | **Details** | |---------------------|------------------------------------------------------------------|-------------------------|-------------------------------------------------------------------| | πŸ₯‡ **Falcon-180B** | [Here](https://huggingface.co/tiiuae/falcon-180b) | *pretrained model* | 180B parameters trained on 3,500 billion tokens. | | Falcon-180B-Chat | [Here](https://huggingface.co/tiiuae/falcon-180b-chat) | *chat model* | Falcon-180B finetuned on a mixture of [Ultrachat](https://huggingface.co/datasets/stingning/ultrachat), [Platypus](https://huggingface.co/datasets/garage-bAInd/Open-Platypus) and [Airoboros](https://huggingface.co/datasets/jondurbin/airoboros-2.1). | | πŸ₯ˆ **Falcon-40B** | [Here](https://huggingface.co/tiiuae/falcon-40b) | *pretrained model* | 40B parameters trained on 1,000 billion tokens. | | Falcon-40B-Instruct | [Here](https://huggingface.co/tiiuae/falcon-40b-instruct) | *instruction/chat model* | Falcon-40B finetuned on the [Baize](https://github.com/project-baize/baize-chatbot) dataset. | | πŸ₯‰ **Falcon-7B** | [Here](https://huggingface.co/tiiuae/falcon-7b) | *pretrained model* | 6.7B parameters trained on 1,500 billion tokens. | | Falcon-7B-Instruct | [Here](https://huggingface.co/tiiuae/falcon-7b-instruct) | *instruction/chat model* | Falcon-7B finetuned on the [Baize](https://github.com/project-baize/baize-chatbot), [GPT4All](https://github.com/nomic-ai/gpt4all), and [GPTeacher](https://github.com/teknium1/GPTeacher) datasets. | | πŸ“€ **RefinedWeb** | [Here](https://huggingface.co/datasets/tiiuae/falcon-refinedweb) | *pretraining web dataset* | ~600 billion "high-quality" tokens. | | Falcon-RW-1B | [Here](https://huggingface.co/tiiuae/falcon-rw-1b) | *pretrained model* | 1.3B parameters trained on 350 billion tokens. | | Falcon-RW-7B | [Here](https://huggingface.co/tiiuae/falcon-rw-7b) | *pretrained model* | 7.5B parameters trained on 350 billion tokens. | # About us The [Technology Innovation Institute](https://www.tii.ae) (TII) is a leading global research center dedicated to pushing the frontiers of knowledge. Our teams of scientists, researchers and engineers work in an open, flexible and agile environment to deliver discovery science and transformative technologies. Our work means we will not only prepare for the future; we will create it. Working together, we are committed to inspiring innovation for a better tomorrow. We are part of Abu Dhabi Government’s Advanced Technology Research Council, which oversees technology research in the emirate. As a disruptor in science, we are setting new standards and serve as a catalyst for change. Faced with a future of limitless possibilities and supported by strategically funded investments, we are encouraging a culture of discovery. Our work reinforces Abu Dhabi and the UAE’s status as an R&D hub and a global leader in breakthrough technologies.