Foundation Models
Foundation models can be applied across domains and tasks. But there are challenges to scalability, and how AI is applied in specific use cases. At IBM Research, we create new foundation models for business, integrating deep domain expertise, a focus on responsible AI, and a commitment to open-source innovation.
Overview
Modern AI models can learn from millions of examples to help find new solutions to difficult problems. But building new systems tends to take time — and lots of data. The next wave in AI will replace task-specific models with ones that are trained on a broad set of unlabeled data that can be used for different tasks — with minimal fine-tuning. These are called foundation models. They can be the foundation for many applications of the AI model. Using self-supervised learning and fine-tuning, the model can apply information it has learned in general to a specific task.
We believe that foundation models will dramatically accelerate AI adoption in business. Reducing time spent labeling data and programming models will make it much easier for businesses to dive in, allowing more companies to deploy AI in a wider range of mission-critical situations. Our goal is to bring the power of foundation models to every enterprise in a frictionless hybrid-cloud environment.
Our work
- Case studyPeter Hess
Simplifying geospatial AI with TerraTorch 1.0
Technical noteRomeo Kienzler, Juan Bernabé-Moreno, Paolo Fraccaro, Bianca Zadrozny, Campbell Watson, Benedikt Blumenstiel, Joao Lucas de Sousa Almeida, Michael Johnston, Christian Pinto, and Michal MuszynskiMeet IBM’s new family of AI models for materials discovery
NewsKim MartineauPhotos: How IBM and NASA's new geospatial model is changing our view of the world
NewsKim MartineauAn IBM-led team is exploring how AI can prepare the electrical grid for the low-carbon era
NewsPeter HessIBM Granite has new experimental features for developers to test
NewsKim Martineau- See more of our work on Foundation Models
Publications
Compress then Serve: Serving Thousands of LoRA Adapters with Little Overhead
- Rickard Gabrielsson
- Jiacheng Zhu
- et al.
- 2025
- ICML 2025
Invariance Makes LLM Unlearning Resilient Even to Unanticipated Downstream Fine-Tuning
- Changsheng Wang
- Yihua Zhang
- et al.
- 2025
- ICML 2025
SPRI: Aligning Large Language Models with Context-Situated Principles
- Hongli Zhan
- Muneeza Azmat
- et al.
- 2025
- ICML 2025
Aligning foundation models on encoded synthetic omic data for patient stratification
- 2025
- ICDH 2025
Multi-View Mixture-of-Experts for predicting molecular properties using SMILES, SELFIES, and graph-based representations
- 2025
- Machine Learning: Science and Tech.
EarthDial: Turning Multi-sensory Earth Observations to Interactive Dialogues
- Sagar Soni
- Akshay Dudhane
- et al.
- 2025
- CVPR 2025
Tools + code
Download Granite on Hugging Face
Explore our family of language, code, time series, and geospatial models.
Download modelsTry Granite for Free
Chat with a Granite model and learn how it can be used across a variety of applications.
View projectRead Granite Documentation
Learn how to access, run, and start using the Granite family of AI models.
View project