IBM Granite Switch 4.1.1: Updating AI Models Like Software

Software teams ship patch releases every week, but language models usually get updated in big, expensive jumps. IBM's Granite team wants to change that. On October 29, 2026, it published Granite Switch 4.1.1, a patch release that adds new capabilities to its modular models without touching the base model at all, according to the team's post on Hugging Face.

• Release type: fully backwards compatible with Granite Switch 4.1, with no migration required
• New speech model: updated Granite Speech recognition covering 10 languages, with native punctuation and capitalization
• Better decisions: an updated Decision adapter in granitelib-core with improved benchmark performance
• What stayed fixed: the base model, which IBM says was not recomputed for this update

What Is Granite Switch?

Granite Switch, which reached version 4.1 on October 15, packages a base language model together with small specialized adapters in a single checkpoint. The 4.1 release came in 8B and 30B sizes and bundled a general Decision adapter, a Guardian adapter for safety-related classification, and the Granite Speech 5.0 Turbo CTC transcription model.

The key engineering idea is the activated LoRA, or aLoRA. Ordinary low-rank adapters change how the model processes the whole prompt, so switching tasks usually means recomputing the key-value cache. IBM built its adapters so the cache generated by the base model can be reused when an adapter is switched on. In practice, one endpoint can answer a chat request, make a classification decision and transcribe audio in quick succession, while using roughly the same GPU memory as the base model alone.

What Changed in Granite Switch 4.1.1?

Two components were refreshed:

• Granite Speech ASR: multilingual speech-to-text across 10 languages, with punctuation and capitalization built in, so transcripts can feed downstream models without a separate formatting pass
• Decision adapter: an updated version in granitelib-core that IBM reports scores better on its decision benchmark

Everything else is identical. As the team puts it, apart from training the new speech model and adapter, nothing was recomputed.

Why Patch-Style Model Updates Matter

For organizations running open-weight models in production, this is the most interesting part of the release. Swapping a whole model means re-running evaluations, re-certifying safety behavior and sometimes re-tuning prompts. If the base model stays frozen and only a targeted adapter changes, the surface area to re-test shrinks dramatically. That is the same logic that makes semantic versioning work in software: a patch number should promise that nothing you depend on has moved.

IBM says it is preparing tutorials and an Adapter SDK so developers and third parties can build, version and ship their own adapters. Granite Switch already runs in vLLM through a plug-in and has native support in llama.cpp, so the path from experiment to deployment is short. You can try it in the team's Speech-to-Decisions demo space on Hugging Face.

This modular approach extends IBM's recent run of open releases, including the Granite Speech 5 open ASR models and the Granite 4.2 reasoning models. Follow the rest of our AI coverage for more open model news.

Sources: IBM Granite on Hugging Face: Granite Switch 4.1.1 — October 29, 2026; IBM Granite on Hugging Face: Granite Switch 4.1 — October 15, 2026; Granite Switch 4.1 30B model card — accessed October 30, 2026.