What changes when a language model gets a patch release like software?

The architecture that makes patches possible

Granite Switch is a modular system. A single base model stays fixed while specialized adapters plug into it. Each adapter adds a distinct capability — speech recognition, decision making, chat — without changing the underlying weights. The 4.1 release demonstrated this pattern. The 4.1.1 release proves it works for incremental updates too.

Nothing was recomputed. The base model remains exactly as it was.

What the new adapters actually do

The updated Granite Speech ASR model handles multilingual speech-to-text across 10 languages. It produces native punctuation and capitalization. Transcripts arrive ready for downstream models without a formatting pass. The model identifier is ibm-granite/granite-speech-5.0-470m-turboctc. It weighs 0.5B parameters. Downloads stand at 61.3k. Likes sit at 111.

The Decision adapter inside granitelib-core shows improved overall performance on the Jev Decision Index test set. The core library identifier is ibm-granite/granitelib-core-r1.0. Downloads reach 5.27k. Likes reach 42. The main Switch model identifier is ibm-granite/granite-switch-4.1-30b. It carries 30B parameters. Downloads show 2.76k. Likes show 42.

How the release happened without retraining the base

Training the new ASR model and Decision adapter was the only compute spent. The base model stayed frozen. This is the point. In traditional model releases a version bump often means a full retrain. Here the version bump means new adapters composed on the same checkpoint. The team released the update just over a week after 4.1. Speed and efficiency are the story.

A live demo space called Speech to Decisions with Granite runs the new stack. It shows live speech heard and sized up by one Granite Switch model. The space reports 67 running instances. The Granite Switch Collection groups five items. It was updated 14 days ago. It has 18 followers.

Where the SDK fits in

Tutorials and an Adapter SDK are in development. The goal is to let developers and third-party providers build their own adapters. They would iterate on versions, test features, and ship updates the way they ship software. No timeline appears in the announcement. The base model stays untouched. The adapters stay swappable. The pattern stays open.