The details
IBM says its Guardian adapter beats several rival decision models on average at safety-related tasks, and the bundled speech model is the fastest on the OpenASR leaderboard. Developers can also build their own Switch models by mixing adapters from IBM’s Granite Libraries, and IBM plans an SDK for training new ones.
Why it matters
Many AI apps chain a chat model, a classifier and a transcription model, each on its own hardware. Packing them into one model can cut cost and delay for companies running AI in production. It also strengthens the case for small, specialised adapters alongside ever-larger general models.