Latest in AI
IBM Granite Switch 4.1.1: What Changed in Speech and Decisions
IBM updates speech recognition and decision adapters without retraining the base model, putting modular model maintenance to a practical test.
IBM announced Granite Switch 4.1.1 on October 29, updating its speech recognizer and decision adapter while leaving the base language model unchanged. The release adds multilingual transcription and improves a decision benchmark, with a more interesting implication for teams that need to maintain models already in production.
According to IBM's announcement, 4.1.1 is backward-compatible with Switch 4.1 and requires no migration. IBM trained the replacement speech component and Decision adapter without recomputing the rest of the model. The company is applying the idea of a software patch to a composed AI system.
What changed in Granite Switch 4.1.1
The updated speech component supports transcription across ten languages, including punctuation and capitalization. That removes a formatting step for applications that otherwise need to clean up raw transcripts before using them downstream. This is an update to the Switch configuration, not the first release of the underlying Granite Speech 5 family.
The Decision adapter also improves on IBM's reported Jev Decision Index results. The announcement describes overall progress, while the adapter documentation provides the version comparison:
| aLoRA adapter | Previous result | Updated result |
|---|---|---|
| Granite 4.1 8B | 42.56 | 46.72 |
| Granite 4.1 30B | 50.91 | 51.33 |
These are IBM-reported benchmark scores, not percentages of tasks your application will necessarily complete correctly. The improvement is also uneven: the smaller adapter moves considerably more than the larger one. There is no basis here for claiming a similar gain in ordinary chat quality.
How the Decision adapter works
Decision takes text or JSON plus a question with a fixed set of answers. It reads probabilities for those answers from next-token logits in a single forward pass rather than writing an explanation and then parsing it. Supported tasks include selecting an option, answering a binary question and scoring an ordinal scale.
The aLoRA variant activates at a marker after the shared state has been processed. That lets questions reuse the base model's cached representation of the state. The adapter is attached to the frozen model at inference time; it does not replace the model's conversational role.
For a support workflow, that creates a useful separation. A transcript can be produced, then classified against a defined routing taxonomy. Updating the transcription component should not require replacing the classifier and language model together. That is an architectural advantage, though the complete workflow still needs regression tests.
What teams should verify before upgrading
Backward compatibility does not mean identical outputs. A changed transcript can alter a downstream decision even when the API and base weights remain unchanged. Test the languages, noisy recordings and routing categories that matter to the actual application, including examples near existing decision thresholds.
The Granite Core Library publishes these adapters under Apache 2.0 and provides artifact-signing information. Pinning component revisions gives an evaluation result a reproducible target and makes a rollback easier to reason about.
IBM says tutorials and an Adapter SDK are still in development. This release should not be read as general availability of a third-party adapter development platform.
The useful next test is whether teams can upgrade one capability, measure its effect on the whole workflow and roll it back independently. Modular releases become valuable when that maintenance process is dependable, not simply when the version number is smaller.
I’m Paras. I work at the intersection of AI, engineering and distribution.
I help technical founders turn model updates and engineering decisions into clear launch stories that developers can understand and share.