IBM released Granite Switch 4.1.1, a recent update to its modular AI model framework. The update, detailed on Hugging Face, focuses on incremental improvements to existing components rather than introducing entirely new capabilities. The primary changes include an updated decision adapter and enhancements to its multilingual speech-to-text model.
The decision adapter within the Granite Switch 4.1.1 release shows improved performance on the Jev Decision Index test set. This adapter is designed for general-purpose decision-making tasks. Earlier versions of Granite Switch, such as 4.1, also featured decision adapters, including a general-purpose one and a specialized Guardian adapter for safety-related decisions. The Granite Switch architecture allows these adapters to be composed with a base model, enabling specialized functions without retraining the entire model.
Alongside the updated decision adapter, Granite Switch 4.1.1 also features an updated Granite Speech ASR model. This speech-to-text model now supports transcription across ten languages and includes native punctuation and capitalization. This integration aims to provide transcripts ready for immediate use by downstream AI models. The original Granite Switch 4.1 release, announced in October 2026, had already integrated a speech-to-text encoder.
The Granite Switch architecture, first introduced in spring 2026, is designed to allow AI models to be updated and configured similarly to software. This modular approach uses "Activated LoRAs" (aLoRAs), which are lightweight adapters that attach to a base model. These adapters can be composed into custom Switch models, allowing for specialized expertise to be added without recomputing the entire model. IBM has stated that the speed and efficiency with which these updates were developed for 4.1.1, without recomputing the base model, highlight the benefits of this architecture.
IBM plans to further support this modular development pattern by releasing tutorials and an Adapter SDK. This will enable developers and third-party providers to create their own adapters, iterate on versions, and deploy updates in a manner akin to software development cycles. The company has previously released Granite Libraries, a repository of specialized adapters for decision-making and other tasks, allowing users to build custom Switch models.
The Granite Switch 4.1.1 release is positioned as a demonstration of the modularity and update efficiency of the Switch architecture. The company highlighted that the base model remains untouched in this update, with only the adapters being modified or improved. This approach contrasts with traditional model development, where updates often require significant retraining.