Most model releases are a full stop and a restart: a new checkpoint, new weights, a fresh API version, a migration guide. IBM’s Granite Switch 4.1.1, published on 29 October, is the counter-example — it is a patch release, and the version number means what it would mean in software.
What actually changed
Granite Switch 4.1 landed a little over a week ago, built on a modular Switch architecture: a single endpoint hosts a base model plus specialized, cache-efficient adapters that can be composed to add narrow expertise. Version 4.1.1 is fully backwards compatible with 4.1 — no migration, incremental improvements under the hood.
Two concrete updates ship in it:
- An updated Granite Speech ASR model (granite-speech-5.0-470m-turboctc), doing multilingual speech-to-text across 10 languages with native punctuation and capitalization, so transcripts arrive pre-formatted for downstream models.
- An updated Decision adapter in granitelib-core, with improved overall performance on the Jev Decision Index test set.
The numbers are unremarkable in isolation. What is remarkable is the cost structure. Apart from training the new ASR model and the Decision adapter, nothing was recomputed — the base model was not touched at all. You compose the adapters you want and deploy.
Why the versioning matters more than the features
There is a strategic claim hidden in the patch number. If the base model is stable and the changing parts are adapters, then model behavior becomes a dependency you version rather than a monolith you re-download. That implies diffing, regression testing, rolling back a single adapter, and shipping a hotfix without re-running the whole training stack. IBM says it is working on tutorials and an Adapter SDK so third-party providers can build, version and ship their own adapters the same way they ship software.
Two caveats are worth stating plainly. First, a patch release is only a patch release if the ecosystem actually treats adapters as separately versioned artifacts — the SDK is not out yet, so today that discipline is IBM’s own. Second, “no migration” is a claim about the interface, not about behavior: a changed Decision adapter can still shift downstream outputs, so the promised regression-testing discipline has to come with it.
The interesting part is not the 470M ASR checkpoint. It is that a frontier-adjacent lab is now describing its model line in semver terms — which suggests the next stage of model delivery is version control, not bigger checkpoints.
Related: Granite Switch 4.1.1 on Hugging Face, ibm-granite/granite-switch-4.1-30b