Introducing Granite Libraries and Project Granite Switch
| Source: IBM Research
Tags: IBM Research, Granite, LLM, model adaptation, parameter-efficient fine-tuning, enterprise AI, modular AI
IBM Research introduces Granite Libraries — plug-and-play adapter functions that customize LLM behavior for specific tasks without retraining the whole model — plus Project Granite Switch, applying software engineering modularity to AI systems.
Details
IBM Research's initiative, framed as 'generative computing,' treats LLMs like modular software stacks rather than monolithic weight files. The core unit is an adapter function: a small trained adapter with a defined input and output that executes a specific task — analogous to a function call in a software library. Granite Libraries is the collection of these adapters; Project Granite Switch is the broader initiative enabling this architecture. The motivation is practical: changing model behavior today requires either full retraining or elaborate prompt engineering — neither scales for enterprise teams needing rapid iteration. IBM's approach separates capabilities behind interfaces, letting different teams develop, test, and swap specific behaviors independently without touching the rest of the model. 'Models are just code with data, just a lot more data than code,' said Luis Lastras, IBM Research's director of language and multimodal models. 'We haven't learned the lessons of software for LLMs — we can build pieces separately.' The framing connects directly to parameter-efficient fine-tuning methods (LoRA, adapters) that are already common in research, but IBM is positioning this as an engineering discipline rather than a research technique. The announcement is conceptually strong but light on implementation specifics — whether this delivers the promised composability in production depends on details not yet disclosed.