Granite Switch 4.1.1 Updates Language Models Like Software

IBM has released Granite Switch 4.1.1, a backward-compatible patch update to its modular language model architecture that lets developers add specialized capabilities through lightweight adapters rather than retraining or recomputing a base model. The update follows Granite Switch 4.1, introduced just over a week earlier, and preserves full compatibility with the prior version while delivering incremental improvements. The release includes an updated Granite Speech automatic speech recognition model that supports multilingual speech-to-text across 10 languages and generates native punctuation and capitalization, producing transcripts ready for downstream processing without additional formatting. IBM also updated the Decision adapter within granitelib-core, noting improved performance on the Jev Decision Index test set. A key aspect of the release is operational efficiency: because the base model remains unchanged, IBM only needed to train and deploy the new ASR model and Decision adapter. The company says this pattern could extend beyond IBM-managed patch releases, as it is preparing tutorials and an Adapter SDK to enable developers and third-party providers to create, version, test, and distribute their own adapters in a software-like delivery model. The new capabilities are already powering IBM’s Speech-to-Decisions demonstration.