Granite Switch 4.1.1 — Update models like you update software

IBM has published Granite Switch 4.1.1, a patch release that follows Granite Switch 4.1 by just over a week. The version numbering is deliberate: 4.1.1 is described as fully backward compatible with 4.1, requiring no migration and bringing incremental improvements under the hood rather than architectural change.
The release carries two substantive updates. The first is an updated Granite Speech ASR model, which handles multilingual speech-to-text across 10 languages and produces native punctuation and capitalization. That matters practically because transcripts come out ready for downstream models without a separate formatting or cleanup pass. The second is an updated Decision adapter in granitelib-core, which the release says performs better overall on the Jev Decision Index test set. The post does not give specific scores or per-language accuracy figures.
What IBM highlights is not the individual features but the effort behind them. Apart from training the new ASR model and the Decision adapter, nothing was recomputed: the base model is untouched. Developers compose the adapters they want and deploy. This is the payoff of the modular Switch architecture introduced in 4.1, which lets a language model be configured with specialized, cache-efficient adapters so new expertise can be added on a single endpoint.
The underlying idea is familiar from software packaging. Rather than retraining or re-releasing a monolithic model each time a capability improves, the base stays fixed and adapters are swapped in like dependency upgrades. Adapters are small trained weight sets layered onto a shared backbone; because they are compact, they can be cached efficiently and served alongside the base without duplicating the full model per capability. The patch-release framing follows from that: if the base is unchanged, a new adapter version should drop in without breaking what already works.
IBM says it is working on tutorials and an Adapter SDK so developers and third-party providers can build their own adapters, iterating on versions and shipping updates the way they ship software. No date is given for either. The company also points readers to its Speech-to-Decisions demo to try the new multilingual and punctuation capabilities.
A likely trade-off is that modularity shifts complexity to composition: users must pick and version the right adapter set, which is easier in a managed single-endpoint setup than if adapters must be tracked and deployed per environment. It also tends to constrain how far a single adapter can diverge from the base before retraining becomes necessary, though IBM has not discussed such limits here.
Why it matters: teams building on Granite Switch can treat speech and decision improvements as routine upgrades rather than migrations, which lowers the cost of staying current. The announced Adapter SDK, if it arrives as described, would let outside providers publish their own adapters on the same versioning pattern, making the ecosystem around the base model more software-like. Developers evaluating the speech update should check its accuracy across their own languages, since no per-language benchmarks were published.
Based on reporting from the original publisher. Visit the source for full context and later updates.