Granite Switch 4.1.1 enables you to update models like you update software.

Meta description: Granite Switch 4.1.1 is a backwards-compatible IBM patch release adding multilingual speech-to-text and an improved Decision adapter for developers.

Tags: Granite Switch 4.1.1, IBM, speech-to-text, Decision adapter, Speech-to-Decisions

Featured image: abstract illustration of modular adapters connecting a speech waveform block and a decision block, friendly flat style, no real people, no logos, no text

IBM has made Granite Switch 4.1.1 available as a patch to Granite Switch 4.1, which launched just over a week ago. For developers working with IBM's modular Switch architecture, the update keeps the existing model endpoint while adding specialized, cache-efficient adapters: an updated Granite Speech ASR model and an improved Decision adapter in granitelib-core. The version number carries the usual software meaning here: no migration is required, and the changes are incremental under the hood.

What the patch adds

The Granite Speech ASR model, a speech-to-text model, is the most concrete user-facing change. It now performs multilingual speech-to-text across 10 languages and produces transcripts with native punctuation and capitalization. That means a voice transcript can be passed to downstream models without a separate formatting pass. For applications that turn spoken input into structured work, the improvement removes a repetitive cleanup step before the next model or workflow sees the text.

The Decision adapter in granitelib-core also received an update. The release notes describe improved overall performance on the Jev Decision Index test set. The claim is qualitative, not a new benchmark number: the patch says performance improved on that test set, while the base model itself was not recomputed.

Why the release speed matters

The release notes highlight speed and efficiency as the broader signal. Apart from training the new ASR model and Decision adapter, nothing was recomputed. The base model stayed untouched, and the team composed the latest adapters and deployed them. That workflow treats model components like software modules: choose the adapter versions you need, combine them, and ship the update without rebuilding the whole language model from scratch.

This is useful for teams that want frequent capability updates without changing the underlying model. A backwards-compatible patch means users can move from 4.1 to 4.1.1 without migration, while still gaining updated speech handling and decision adapter behavior. The update also shows how the architecture can add new expertise on a single endpoint rather than requiring a new model family for each capability.

What comes next

The Adapter SDK and tutorials are the part to watch. IBM says it is working on them so developers and third-party providers can build their own adapters, iterate on versions, test features, and ship updates the way they ship software. That points toward a more extensible model: new expertise would not be limited to patch releases to an IBM-provided model. The release notes say more details are coming soon, so the SDK should not be treated as available now.

What to try next

For now, the update points users toward the latest Granite Switch models and the Speech-to-Decisions demo. The demo is powered by the new multilingual and punctuation capabilities, so it is a natural place to see how updated speech transcription and adapter-based composition work together. If you are evaluating Granite Switch 4.1, the 4.1.1 patch offers the same endpoint model with updated components and no migration step. If you are building voice workflows, focus on whether the updated Granite Speech ASR model reduces the need for transcript formatting before downstream models.