IBM shipped Granite Switch 4.1.1 on 29 October as a patch release that retrains nothing except two plug-in adapters, leaving the base language model completely untouched S¹. The approach treats a large model less like a monolith you rebuild from scratch and more like a codebase you hot-swap components in.
My read: This is the first model architecture I've seen where a version bump genuinely resembles a software patch rather than a full retrain. The "fully backwards compatible, no migration" claim is a vendor assertion nobody outside IBM has verified, and the Decision adapter improvement ships without published scores. But the structural argument, that you can upgrade speech recognition and decision-making independently while the base stays frozen, is the right bet if adapter swapping holds up at production scale.
The Switch architecture, introduced in version 4.1, runs a base model at a single endpoint alongside specialised adapters that cache efficiently and inject new capabilities S¹. Think of a fixed engine with swappable cylinder heads. The GitHub repository, generative-computing/granite-switch, carries an Apache 2.0 licence and has drawn 115 stars since its creation in May 2026 P³.
Two adapters, one frozen base
IBM's updated Granite Speech ASR adapter transcribes spoken audio in ten languages and inserts punctuation and capital letters automatically, so the resulting text can feed directly into other models rather than requiring a cleanup step first S¹. The specific languages are not named in the release, so anyone evaluating the model for a particular market would need to test it directly.
The Decision adapter inside granitelib-core scores higher on the Jev Decision Index benchmark, which IBM describes as "improved overall performance" S¹. IBM has not published scores or error bars, so the degree of improvement is unverified. Beyond the two adapter updates, every other component matches 4.1 — no part of the base model was recomputed or altered S¹. IBM states on its Hugging Face blog that existing 4.1 deployments can move to 4.1.1 without any migration work, though no outside party has validated that claim S¹.
Who this reaches and what is missing
A team running Granite Switch 4.1 in production for call-centre transcription could pull the 4.1.1 revision and gain multilingual ASR with built-in punctuation without redeploying the base model or retesting their existing chat and decision workflows.
The upgrade surface shrinks from the entire model to two adapters.
A Hugging Face demo labelled Speech-to-Decisions lets developers try the updated pipeline from audio input through to model output S¹, so a developer can check whether the 10-language coverage fits their market before committing. IBM's adapter-swapping approach competes on a different axis from the broader open-weight ecosystem: not how many models you can download, but how cheaply you can upgrade the ones you already run.
IBM is developing an Adapter SDK along with tutorial materials that would let external developers and vendors create custom adapters for the architecture S¹. Both are in-development commitments, not released features. Granite Switch is trying to solve a problem one level below orchestration tools that manage multiple models: updating a single model without touching its core.
The next checkpoint is the Adapter SDK, which has no published release date S¹.
Sources: S1 — Granite Switch 4.1.1 — Update models like you update software · P2 — ibm-granite/granite-switch-4.1-30b · granite-switch 4.1.1: updated dec · P3 — generative-computing/granite-switch · P4 — Update README.md · ibm-granite/granite-switch-4.1-8b at f94f440 · P5 — ibm-granite/granite-speech-models
Written from 5 sourced items, 5 of them primary.
