A point release for a model
IBM published Granite Switch 4.1.1 on 29 October, and the framing is the news: the company presents it as a patch release in the ordinary software sense — fully backwards compatible with 4.1, no migration, incremental improvements underneath. The post’s title says it directly: update models like you update software.
Two things changed. A new speech recognition model, granite-speech-5.0-470m-turboctc at 0.5B parameters, replaces the previous one. And the Decision adapter in granitelib-core was updated. The 30B base model was not retrained, and IBM says nothing else was recomputed. That is what makes the version number defensible: a patch that touches two components and leaves the weights everyone depends on alone.
The claim with no number attached
IBM says the updated Decision adapter has “improved overall performance on the Jev Decision Index test set.” The post gives no figure — not a before, not an after, not a delta. For a release whose entire argument is that model updates should behave like software updates, that is the wrong omission. A dependency bump that says it is better without saying by how much is the thing careful teams refuse to apply.
The speech model gets the same treatment. The post says it supports 10 languages with native punctuation and capitalisation, and reports no word error rate or comparison against the model it replaces.
Ten languages, or seven
The blog post and the model card do not agree. The post says 10 speech languages. The revision table on the granite-switch-4.1-30b model card, where 4.1.1 is listed as the latest release on main, puts the Speech column at “7 languages”. The same table lists 4.1’s Speech column as “English” — so the jump is real either way, but its size depends on which IBM page you read. Both were current on the same day.
The model card also states a licence, Apache 2.0, which the blog post does not mention at all.
Why the versioning idea is worth taking seriously anyway
The underlying proposal is sound and unusual. Model releases normally arrive as whole new artefacts with new names, new behaviour and no compatibility promise, which forces every downstream team into a re-evaluation. Semantic versioning with a stable base and swappable adapters would let a team take a speech improvement without re-testing everything that touches the language model.
IBM says it is working on tutorials and an Adapter SDK so that developers and third-party providers can version and ship their own adapters, with no date beyond “more soon.” That is the piece that would make the idea more than a naming convention.
What to watch
Three things would settle whether this is versioning or branding. Published numbers for the Decision adapter claim. A single authoritative statement of how many languages the speech model handles. And the Adapter SDK, which is the only part of the announcement that would let anyone outside IBM test whether a model really can be patched like a dependency.
