Institute of Foundation Models
Message from the President
The mission of the IFM is to position our academic research at the forefront of Generative AI development, and ground it on substantial societal needs. This institute will be the hub of activity for fundamental research in areas we already have pioneered with the ability to focus those energies now into a wide spectrum of powerful foundation models such as the GPTs for lingual, visual, and biological data that can be used in industries dedicated to healthcare, climate change and sustainability, and in turn advance the innovation of foundation models itself. The options are endless and we look forward to partnering with others in this area.
Professor Eric Xing
MBZUAI President
Who We Are
We are a dedicated research hub for building, understanding, using, and risk-managing Foundation Models including the large language models underlying the ChatGPTs, and transformers pre-trained on biological data underlying those tools like AlphaFold2, and many more to emerge. Our mandate is to advance research, nurture the next generation of AI builders, and drive transformative contribution to UAE’s knowledge-driven economy.
Why Foundation Models
Foundation Models stand as a bedrock of advanced AI, enabling versatile and deeper machine insights, transformative content generation capability, and potentially super-human problem-solving skills. We must make this new technology affordable, safe, ethical, standardizable, and certifiable.
Careers within IFM
Our Values
Research
Pursuing advanced studies
to push AI frontiers.
Talent
Cultivating and nurturing
top AI innovators.
Artifacts
A Diverse Array of
Foundational Models.
Empower
Enabling others through
AI capabilities and insights.
Projects
LLM360: Open-source LLMs for Transparency, Trust, and
Collaborative Research
LLM360 is a groundbreaking open-source framework aimed at enhancing
transparency and collaboration in AI research. It offers detailed insights by providing full access to
training processes, including checkpoints, datasets and source code for models like Amber and
CrystalCoder. LLM360 aims to democratize LLMs by offering a framework for open-source LLMs to foster
transparency, trust, and collaborative research.
Vicuna: A “Decarbonized” Vet High-Performance LLM
Vicuna is a lightweight, open-source LLM chatbot fine-tuned with user-shared
conversations from ShareGPT, surpasses 90%* quality of OpenAI ChatGPT and Google Bard, outperforming
models like LLaMA and Stanford Alpaca in over 90%* of cases, all achieved at a training cost of
approximately $300.
GET: A Foundation Model for Genome
GET is a foundation model for transcriptional regulation that accurately
predicts gene expression in diverse human cell types, adapts to different conditions and sequencing
platforms, and identifies regulatory elements and interactions, making it a valuable tool for gene
regulation analysis.
JAIS – World’s Best Arabic LLM
Jais, an open-source Arabic large language model, surpasses existing models in
quality and competes strongly with English models despite limited English data, serving over 400 million
Arabic-speaking individuals worldwide.
ClimateGPT: Our Commitment to the COP28’s Mission on
Sustainability
ClimateGPT is a specialized LLM for conversations related to Climate Change and
Sustainability topics in both English and Arabic languages.
GLaMM: A Multimodal Foundation Model with Pixel Grounding
Capabilities
The first LLM with pixel-level visual grounding capabilities with the
flexibility to process both image and region inputs.
The “Zoo” of LLMs
Visit our “Zoo” of LLMs where, a diverse ensemble of foundational models are
being developed, forming a versatile and dynamic ecosystem for a wide range of
language/visual/audio-related tasks and applications.
COMPRER: A multi-modal foundation model for cardiovascular
disease
COMPRER is a foundation model for enhancing medical-image representation,
diagnostic inferences, and prognosis of diseases. It integrates high-resolution fundus images and
ultrasound images of the carotid arteries into a single model that encompasses both modalities. The
model achieves higher scores in evaluating medical conditions and predicting future cardiovascular
events compared to existing models.