On October 29, IBM presented Granite Switch 4.1.1 as a compatible patch with updated speech and decision components and an unchanged base model. Its announcement describes better decisions and transcription, but is not an independent performance measurement. This is a follow-up revision to the previously introduced Switch 4.1, rather than another announcement of the original product.
What actually changes
The 30B model card dates the revision to October 27. It retains the guardian-core and decision adapters, updating the latter's weights. Control tokens select an adapter; documentation says its existing invocation format remains unchanged. Distinguishing the artifact date from the later announcement matters for teams tracking releases.
The underlying idea is to narrow the scope of a change. If an application uses separate specialisations for conversation and evaluation, an evaluation update can be considered separately from other behaviour. That is our editorial explanation of modularity, not evidence that every combination is risk-free.
There is a concrete discrepancy worth preserving. The blog describes ten languages for the updated speech recogniser. The card and revision table for this 30B composition list seven: English, French, German, Spanish, Portuguese, Italian and Japanese. We therefore do not treat ten as established support for this artifact, or infer that Bulgarian is included.
The version belongs in the result
The repository explains that main follows the latest revision while releases have tags. BUILD.md documents the composition. For a reproducible trial, this provides a useful starting point: record the specific artifact instead of only its family name. Otherwise, two runs may appear identically labelled while using different components.
Our suggested evaluation would retain representative tasks for each adapter actually used. In a conversation-processing application, for example, a team could separately examine the transcription, selected category and application response to an ambiguous utterance. An improvement at one stage would then not automatically count as an improvement to the entire workflow.
For speech, checking the team's own recordings, noise conditions and punctuation would be useful. Better formatting can help downstream processing, but correct commas do not establish that a name or number was recognised correctly. Evaluation of numeric and named fields should keep those errors visible.
This also suggests a meaningful comparison between revisions: hold the input set and application rules fixed, then inspect changed outputs. A new version number alone cannot explain whether a difference came from recognition, classification or the surrounding application. Such a comparison is an evaluation proposal; SIExpo has not run this model locally.
What remains forthcoming
IBM describes an Adapter SDK and tutorials for custom adapters as work in progress. We do not present them as a completed public capability. We also have no independent measurement of training savings, serving cost or accuracy on Bulgarian data.
The editorial conclusion is specific: this patch makes the management of individual AI components a more visible engineering concern. Its practical value must be established for the exact version and task. Interface compatibility helps integration, while checking the resulting behaviour remains the deploying team's responsibility.
*Cover: AI illustration of a modular update, not a depiction of actual hardware.*
*Editorial verification: 31.10.2026, 09:12:27 +0200 (Europe/Sofia).*
