IBM Granite Switch 4.1.1 patch release adds multilingual ASR, improved decision adapter

IBM released Granite Switch 4.1.1, a patch-level update to its modular large language model architecture that maintains full backward compatibility with version 4.1. The release updates two specialized adapters: the Granite Speech ASR model now supports multilingual speech-to-text across ten languages with native punctuation and capitalization, and the Decision adapter in granitelib-core shows improved performance on the Jev Decision Index benchmark. The core base model remains unchanged; only the adapters were recomputed and composed, demonstrating the architecture's design goal of enabling rapid, software-like iteration cycles for AI model components.

IBM emphasized that this release pattern is intended to extend beyond IBM's own models. The company is developing tutorials and an Adapter SDK to allow independent developers and third-party providers to build, version, test, and deploy their own adapters using the same compositional approach. This positions the Switch architecture as a potential open platform for modular AI development rather than a closed vendor system.

Adapter architecture and update mechanics

• Modular adapter composition: Granite Switch uses cache-efficient specialized adapters that layer onto a frozen base model, allowing targeted capability updates without full model retraining or migration.
• Granite Speech ASR update: The speech-to-text adapter now handles ten languages with integrated punctuation and capitalization, eliminating the need for separate post-processing formatting before downstream model consumption.
• Decision adapter improvement: The granitelib-core Decision adapter shows measurable gains on the Jev Decision Index test set, a benchmark for structured decision-making performance.
• Backward compatibility guarantee: Version 4.1.1 requires no migration from 4.1, with the version semantics matching software patch conventions rather than typical ML model release practices.
• Future Adapter SDK: IBM is building tooling and documentation to enable third-party adapter development, though no release date or license specifics for the SDK were provided in this announcement.

Why it matters

The adapter-based update model addresses a critical pain point in production AI deployments: the cost and operational disruption of full model replacements. For Linux and open-source AI infrastructure, compositional architectures that separate base weights from task-specific adapters could reduce storage, memory, and serving overhead while enabling faster security and capability patches. The promised Adapter SDK will be particularly significant if released under a permissive open-source license, as it would allow the broader community to build interoperable extensions rather than depending solely on IBM's release cycle.