The details
Granite Switch, released just over a week earlier, lets one model endpoint load specialised add-ons called adapters as needed. This update is fully backwards compatible, with no migration needed. The new speech model adds punctuation and capitalisation, so transcripts are ready for other models without a cleanup step. The decision adapter scores better on IBM’s Jev Decision Index test set. Only those two parts were retrained; everything else was left as it was. IBM says it is preparing tutorials and an Adapter SDK so outside developers can build and update their own adapters.
Why it matters
Updating an AI model usually means retraining or swapping the whole thing. If companies can patch one skill at a time, they can ship fixes faster and test changes with less risk, much as they update software. The approach will matter more if other developers start building adapters with IBM’s promised toolkit.