Granite Switch 4.1.1: update models like software

What changed

Less than two weeks after Granite Switch 4.1, IBM shipped the 4.1.1 patch release (Hugging Face blog). Switch is a modular language model built from a base model, swappable adapters and a speech encoder, so a single endpoint can chat, make decisions and transcribe audio. The version number means what it does in software: fully backwards compatible with 4.1, no migration needed.

• Updated weights for the decision adapter (aLoRA). Its prompt format, invocation and readout are unchanged.
• The embedded Granite Speech 5.0 TurboCTC encoder (470M parameters) is no longer English only: per the model card it transcribes English, French, German, Spanish, Portuguese, Italian and Japanese, with punctuation and capitalization. The blog post mentions 10 languages, while the model card lists seven.
• The guardian-core safety adapter and the 30B, 128K-context Granite 4.1 base model stay the same.

How good is it?

In IBM's own evaluation, the updated decision adapter scores 51.33 overall on the public Jev Decision Index v0.3 set (a chance-corrected weighted aggregate). By category: knowledge 34.8, language 58.5, retrieval 59.8, tools 69.0, arts 25.9. The model card doesn't list the 4.1 score, so the size of the improvement isn't shown.

Why it matters

IBM says nothing was recomputed apart from training the new ASR model and decision adapter; the base model is untouched. The pattern promises model updates that work like library updates: every release is tagged in the repo, and you can pin the previous one with a revision parameter. IBM is working on tutorials and an Adapter SDK so third parties can build and ship their own adapters.

Try it

granite-switch-4.1-30b is Apache 2.0 licensed and runs on Transformers and vLLM, with an 8B variant also available. Install the granite-switch 0.1.1 package (pip install "granite-switch[vllm,audio]>=0.1.1"). Audio goes through the OpenAI-compatible chat API as an input_audio part, and the language is detected automatically. One limitation: the speech encoder and the LLM aren't jointly trained yet, so the transcript is simply spliced into the chat context.